Weather radar meteorological echo and non-meteorological echo identification system and method

Abstract

The invention discloses a weather radar meteorological echo and non-meteorological echo identification system, and belongs to the technical field of meteorological radar signal processing. The method comprises the following steps: acquiring radar original data containing reflectivity factors, radial velocity and spectral width, and generating a three-dimensional feature matrix through format standardization, combined noise removal, neighborhood interpolation filling and parameter normalization; constructing a sample set through professional labeling, consistency verification and multi-dimensional data enhancement; a deep convolutional neural network containing a ResNet50 feature extraction module, a CBAM double attention module and a classification output module is constructed, and a binary cross entropy loss function, an Adam optimizer and an early stop regularization strategy are adopted for training optimization; to-be-recognized data is preprocessed and then input into the model, and a classification result and confidence are output. The method does not need manual feature design, significantly improves the recognition stability of echoes difficult to distinguish in a complex scene, adapts to different radars and regional environments, and meets the real-time processing requirements.

Description

Technical Field

[0001] This application relates to the field of weather radar signal processing technology, and in particular to a weather radar system and method for identifying weather radar echoes and non-weather echoes. Background Technology

[0002] Weather radar, as a core device for meteorological observation and disaster early warning, transmits electromagnetic waves and receives reflected echoes from targets. This enables precise detection of the location, intensity, and movement of meteorological targets in the atmosphere, providing crucial data support for short-term weather forecasts, severe convective weather warnings, and other meteorological operations. However, in actual radar detection, the echo signals not only include meteorological echoes from rain, snow, and hail, but also often contain non-meteorological echoes from ground features (such as mountains and buildings), superrefractive objects, insects, and ocean waves. These non-meteorological echoes can interfere with the accuracy and reliability of meteorological data, reducing the reliability of weather forecasts and disaster early warnings.

[0003] To distinguish between meteorological and non-meteorological echoes, the industry has developed several traditional identification methods, mainly including: First, the threshold method, which classifies echoes into two categories based on empirical thresholds for parameters such as reflectivity factor and radial velocity. However, the threshold setting of this method is significantly affected by regional characteristics and weather conditions, and its adaptability is poor in complex environments. Second, the morphological method, which distinguishes echoes by their spatial continuity, area, shape, and other morphological features. However, its identification accuracy is difficult to guarantee for non-meteorological echoes with complex shapes. Third, the statistical feature method, which extracts statistical features such as the mean, variance, and texture of the echoes and combines them with traditional machine learning algorithms such as support vector machines (SVM) and decision trees to achieve classification. However, this method relies on manually designed features and is difficult to capture the deep and complex features hidden in the echo data, and so on.

[0004] The aforementioned traditional identification methods generally suffer from the following shortcomings: Identification accuracy is significantly affected by environmental factors, exhibiting insufficient robustness in complex terrains such as mountainous and coastal areas, or in scenarios with multiple weather types overlapping; their adaptability to the diversity of non-meteorological echoes is limited, particularly for difficult-to-distinguish non-meteorological echoes such as super-refracted echoes and faint insect echoes, resulting in poor identification performance; furthermore, the limitations of manually designed features lead to weak model generalization ability, making it difficult to adapt to different types of weather radars; and finally, some methods have complex processing procedures and insufficient real-time performance, failing to meet the processing requirements of real-time data streams from weather radars, etc.

[0005] With the increasing demands for radar data quality in meteorological operations, there is an urgent need for an echo recognition technology that can automatically extract deep features, has high recognition accuracy, strong robustness, and good real-time performance. This technology can solve the problems of traditional methods relying on human experience and having poor adaptability, and provide more reliable data support for weather forecasting and disaster early warning. Summary of the Invention

[0006] To address the technical problem in existing weather radar echo recognition technologies that "rely on manual experience to design features or set thresholds, resulting in low recognition reliability and weak generalization ability in complex scenarios," this invention provides a method for identifying meteorological and non-meteorological echoes from weather radar. The technical solution is as follows: On the one hand, a method for identifying meteorological echoes and non-meteorological echoes from weather radar is provided, including: Raw weather radar echo data containing core parameters such as reflectivity factor, radial velocity, and spectral width are acquired. The raw echo data undergoes format standardization, a combined noise removal process combining adaptive median filtering and wavelet thresholding, invalid value filling via neighborhood interpolation, and parameter normalization to generate a unified three-dimensional echo feature matrix. In the combined noise removal, the adaptive median filtering window size is 3×3~7×7, dynamically adjusted according to noise density. Wavelet thresholding uses the db4 wavelet basis function, and after three-level wavelet decomposition, soft thresholding is applied to high-frequency coefficients. The parameter normalization process maps parameters to the [0,1] interval, with the mapping formula being x... norm =(xx min ) / (x max -x min ), where x is the original parameter value, x min x max The global minimum and maximum values ​​of this parameter in historical data; the search range for invalid value filling by the neighborhood interpolation method is a 3×3~5×5 neighborhood area around the invalid value pixel. The filling is completed by calculating the weighted average of the effective pixels in the neighborhood. The weight decreases linearly with the distance of the pixel from the invalid value point. The three-dimensional echo feature matrix is ​​manually labeled, with the labeling category being either meteorological echo or non-meteorological echo. The labeling results are verified for consistency, and samples with a labeling consistency rate ≥98% are included in the effective sample set. The training set, validation set, and test set are divided in a 7:1:2 ratio. The training set samples are expanded using multi-dimensional data augmentation techniques, including random rotation, horizontal / vertical flipping, Gaussian noise addition, and local cropping. In the multi-dimensional data augmentation techniques, the random rotation angle is 0°, 90°, 180°, or 270°, the variance of the Gaussian noise ranges from 0.01 to 0.05, and the local cropping ratio ranges from 0.8 to 1.0. After data augmentation, the number of training set samples is expanded by 3 to 5 times. A deep convolutional neural network model is constructed, which sequentially includes an input layer, a feature extraction module, an attention enhancement module, and a classification output module. The input layer is used to receive the three-dimensional echo feature matrix. The feature extraction module adopts a ResNet50 residual network and extracts multi-scale features through five convolutional stages and residual connections. The attention enhancement module is a CBAM convolutional block attention module, which assigns weights to the feature map through channel attention sub-modules and spatial attention sub-modules in sequence. The classification output module outputs the binary classification result and confidence score through global average pooling, two fully connected layers, and a sigmoid activation function. The training set is input into the model, and the prediction bias is calculated using the binary cross-entropy loss function. The model parameters are dynamically adjusted using the Adam adaptive momentum optimizer. L2 regularization and early stopping strategies are introduced to suppress overfitting. The key hyperparameters, such as learning rate, kernel size, and batch size, are optimized using a grid search method. The initial learning rate of the Adam adaptive momentum optimizer is 0.001, and the weight decay coefficient is 0.0001. A learning rate decay strategy is adopted, which reduces the learning rate to 0.5 every 50 rounds. The early stopping strategy is triggered when the accuracy of the validation set does not improve for 10 consecutive rounds. The hyperparameter optimization range of the grid search method is: learning rate 0.0001~0.01, kernel size 3×3 or 5×5, batch size 16~64, and training iterations 50~200 rounds. After standardizing the radar echo data to be identified according to the data acquisition and preprocessing steps, the data is input into the trained optimal model, and the meteorological echo / non-meteorological echo classification results and the confidence level in the 0~1 interval are output.

[0007] Optionally, the format standardization conversion transforms the raw data of IRIS and NetCDF formats into a unified three-dimensional matrix of H×W×3, where H and W are the height and width of the echo image, with values ​​ranging from 64 to 256, and the three channels correspond to the reflectivity factor, radial velocity, and spectral width parameters, respectively.

[0008] Optionally, in the combined noise removal, the window size of the adaptive median filter is 3×3~7×7, which is dynamically adjusted according to the noise density; Wavelet thresholding denoising uses the db4 wavelet basis function, and after 3-level wavelet decomposition, soft thresholding is applied to the high-frequency coefficients. The parameter normalization process involves mapping the parameter to the [0,1] interval. The mapping formula is xnorm=(x-xmin) / (xmax-xmin), where x is the original parameter value, and xmin and xmax are the global minimum and maximum values ​​of the parameter in historical data.

[0009] Optionally, the search range for invalid value filling in the neighborhood interpolation method is a 3×3~5×5 neighborhood area around the invalid value pixel. The filling is completed by calculating the weighted average of the effective pixels in the neighborhood, and the weight decreases linearly with the distance of the pixel from the invalid value point.

[0010] Optionally, in the multi-dimensional data augmentation technique, the random rotation angle is 0°, 90°, 180° or 270°, the variance of Gaussian noise is in the range of 0.01~0.05, the local cropping ratio is in the range of 0.8~1.0, and the number of training set samples is increased by 3~5 times after data augmentation.

[0011] Optionally, the channel attention submodule performs global average pooling and global max pooling on the feature map, inputs the two pooling results into a shared fully connected layer, and generates a channel weight vector through Sigmoid activation; the spatial attention submodule performs average pooling and max pooling on the channel-weighted feature map, concatenates the pooling results, and generates a spatial weight map through a convolutional layer and Sigmoid activation.

[0012] Optionally, the Adam adaptive momentum optimizer has an initial learning rate of 0.001 and a weight decay coefficient of 0.0001, and adopts a learning rate decay strategy that decays to 0.5 every 50 rounds; the early stopping strategy is triggered when the validation set accuracy does not improve for 10 consecutive rounds; the hyperparameter optimization range of the grid search method is: learning rate 0.0001~0.01, convolution kernel size 3×3 or 5×5, batch size 16~64, and training iterations 50~200 rounds.

[0013] On the other hand, a weather radar meteorological echo and non-meteorological echo identification system is provided, applicable to the aforementioned weather radar meteorological echo and non-meteorological echo identification method. The system includes: The data acquisition module is used to acquire raw echo data from weather radar containing core parameters such as reflectivity factor, radial velocity, and spectral width. It performs format standardization conversion, combined noise removal using adaptive median filtering and wavelet thresholding, invalid value filling via neighborhood interpolation, and parameter normalization on the raw echo data to generate a unified three-dimensional echo feature matrix. In the combined noise removal, the window size of the adaptive median filtering is 3×3~7×7, dynamically adjusted according to the noise density. Wavelet thresholding uses the db4 wavelet basis function, and after three-level wavelet decomposition, soft thresholding is applied to the high-frequency coefficients. The parameter normalization process maps the parameters to the [0,1] interval, with the mapping formula being x... norm =(xx min ) / (x max -x min ), where x is the original parameter value, x min x maxThe global minimum and maximum values ​​of this parameter in historical data; the search range for invalid value filling by the neighborhood interpolation method is a 3×3~5×5 neighborhood area around the invalid value pixel. The filling is completed by calculating the weighted average of the effective pixels in the neighborhood. The weight decreases linearly with the distance of the pixel from the invalid value point. The sample management module is used to manually annotate the three-dimensional echo feature matrix, with the annotation category being either meteorological echo or non-meteorological echo. The module performs consistency verification on the annotation results, and includes samples with an annotation consistency rate ≥98% into the effective sample set. The training set, validation set, and test set are divided in a 7:1:2 ratio. The training set samples are expanded using multi-dimensional data augmentation techniques, including random rotation, horizontal / vertical flipping, Gaussian noise addition, and local cropping. In these multi-dimensional data augmentation techniques, the random rotation angle is 0°, 90°, 180°, or 270°, the variance of the Gaussian noise ranges from 0.01 to 0.05, and the local cropping ratio ranges from 0.8 to 1.0. After data augmentation, the number of training set samples is increased by 3 to 5 times. The model training module is used to construct a deep convolutional neural network model that sequentially includes an input layer, a feature extraction module, an attention enhancement module, and a classification output module. The input layer is used to receive a three-dimensional echo feature matrix. The feature extraction module uses a ResNet50 residual network and extracts multi-scale features through five convolutional stages and residual connections. The attention enhancement module is a CBAM convolutional block attention module, which assigns weights to the feature map through channel attention sub-modules and spatial attention sub-modules in sequence. The classification output module outputs the binary classification result and confidence score through global average pooling, two fully connected layers, and a Sigmoid activation function. The bias prediction module inputs the training set into the model, calculates the prediction bias using the binary cross-entropy loss function, dynamically adjusts the model parameters using the Adam adaptive momentum optimizer, introduces L2 regularization and an early stopping strategy to suppress overfitting, and optimizes key hyperparameters such as learning rate, kernel size, and batch size using a grid search method. The initial learning rate of the Adam adaptive momentum optimizer is 0.001, the weight decay coefficient is 0.0001, and a learning rate decay strategy of reducing the learning rate to 0.5 every 50 rounds is adopted. The early stopping strategy is triggered when the accuracy of the validation set does not improve for 10 consecutive rounds. The hyperparameter optimization range of the grid search method is: learning rate 0.0001~0.01, kernel size 3×3 or 5×5, batch size 16~64, and training iterations 50~200 rounds. The model output module is used to standardize the radar echo data to be identified according to the data acquisition and preprocessing steps, input it into the trained optimal model, and output the meteorological echo / non-meteorological echo classification results and the confidence level in the 0~1 interval.

[0014] This invention discloses a weather radar echo and non-weather echo identification system and detection method, belonging to the field of weather radar signal processing technology. It collects raw radar data containing reflectivity factors, radial velocity, and spectral width, and generates a three-dimensional feature matrix through format standardization, combined noise removal, neighborhood interpolation filling, and parameter normalization. A sample set is constructed through professional annotation, consistency verification, and multi-dimensional data augmentation. A deep convolutional neural network containing a ResNet50 feature extraction module, a CBAM dual attention module, and a classification output module is built, and trained and optimized using a binary cross-entropy loss function, an Adam optimizer, and an early stopping regularization strategy. The data to be identified is preprocessed and input into the model, outputting classification results and confidence scores. This method eliminates the need for manual feature design, significantly improves the stability of echo identification in complex scenarios, adapts to different radar and geographical environments, and meets real-time processing requirements. Attached Figure Description

[0015] Figure 1 A flowchart illustrating a method for identifying meteorological and non-meteorological echoes from a weather radar. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0017] In this article, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0018] In the actual implementation of this invention, the environment configuration is completed first. The hardware uses an Intel Xeon Gold 6248 CPU, an NVIDIA Tesla V100 GPU (32GB of video memory), 64GB of RAM and a 2TB SSD solid-state drive, equipped with a gigabit Ethernet interface to ensure data flow transmission. The software environment uses the Ubuntu 20.04 LTS operating system, paired with the PyTorch 1.10 deep learning framework, Python 3.8 programming language, supplemented by data processing libraries such as NumPy, Pandas, and OpenCV, visualization libraries such as Matplotlib and TensorBoard, and data format parsing libraries such as PyART and xarray to ensure compatibility and stability of each component.

[0019] Example 1 like Figure 1 The diagram illustrates a flowchart of a method for identifying meteorological and non-meteorological echoes from a weather radar. The method includes: Step 101: Obtain raw weather radar echo data containing core parameters such as reflectivity factor, radial velocity, and spectral width, and generate a three-dimensional echo feature matrix with uniform dimensions.

[0020] Among them, the original echo data is subjected to format standardization conversion, combined noise removal of adaptive median filtering and wavelet threshold denoising, invalid value filling by neighborhood interpolation and parameter normalization.

[0021] In one possible implementation, during the data acquisition phase of step 101, radar data from three different meteorological observation stations in the North China Plain, the Southwest Mountainous Area, and the Southeast Coast of my country are selected, covering both S-band and C-band types, spanning one year, and including different weather types such as sunny, rainy, snowy, and hail. A total of 100,000 frames of raw data are collected, including 60,000 frames in IRIS format and 40,000 frames in NetCDF format. Each frame contains three core parameters: reflectivity factor (Z, in dBZ), radial velocity (Vr, in m / s), and spectral width (W, in m / s). After data acquisition, the data preprocessing stage begins. The data preprocessing module converts the raw data of different formats into a unified 128×128×3 three-dimensional matrix (H and W can be flexibly adjusted between 64 and 256 according to the radar resolution, and the three channels correspond to three types of core parameters respectively). Then, a combination strategy of "adaptive median filtering + wavelet threshold denoising" is used to remove noise. The adaptive median filtering window size is dynamically adjusted between 3×3 and 7×7 according to the noise density, effectively removing impulse noise and preserving edge features. Wavelet threshold denoising uses the db4 wavelet basis function to perform soft thresholding on the high-frequency coefficients after three-level decomposition to eliminate random noise interference. To address invalid values ​​caused by radar detection blind spots or data transmission loss, a 3×3 to 5×5 neighborhood interpolation method is used to fill the gaps. The data is filled by calculating the weighted average of the effective pixels in the neighborhood (the weight decreases linearly with the distance from the invalid value point). Finally, all parameters are mapped to the [0,1] interval according to the mapping formula xnorm=(x-xmin) / (xmax-xmin) to eliminate the difference in magnitude. Here, x is the original parameter value, and xmin and xmax are the global minimum and maximum values ​​of the parameter in the historical data.

[0022] Step 102: Manually annotate the three-dimensional echo feature matrix, labeling it as either meteorological echo or non-meteorological echo, verify the consistency of the annotation results, and expand the training set samples.

[0023] Among them, samples with a labeling consistency rate of ≥98% are included in the effective sample set, and the training set, validation set and test set are divided in a ratio of 7:1:2. The training set samples are expanded by multi-dimensional data augmentation techniques such as random rotation, horizontal / vertical flipping, Gaussian noise addition and local pruning.

[0024] In one possible implementation, during the sample set construction phase of step 102, two engineers with eight years of experience in weather radar data processing independently and manually annotate the preprocessed 100,000-frame feature matrix. The annotation categories are divided into "weather echoes" (including rainfall, snowfall, hail, etc.) and "non-weather echoes" (including ground features, super-refraction, insects, ocean waves, etc.). After the annotation is completed, the consistency verification unit of the sample management module compares the results, and only 85,000 frames with an annotation consistency rate ≥98% are included in the effective sample set, of which 51,000 are weather echoes and 34,000 are non-weather echoes. The dataset was automatically divided into a training set of 59,500 frames, a validation set of 8,500 frames, and a test set of 17,000 frames in a 7:1:2 ratio. To improve the model's generalization ability and address the imbalanced sample problem, multi-dimensional data augmentation operations were performed on the training set: random rotations of 0° / 90° / 180° / 270°, horizontal / vertical flipping, addition of Gaussian noise with a variance of 0.03 (the variance range can be adjusted between 0.01 and 0.05), and local cropping at a ratio of 0.9 (cropping ratio 0.8 to 1.0). Ultimately, the training set sample size was expanded to 238,000 frames.

[0025] Step 103: Construct a deep convolutional neural network model that sequentially includes an input layer, a feature extraction module, an attention enhancement module, and a classification output module.

[0026] The input layer receives the 3D echo feature matrix. The feature extraction module uses a ResNet50 residual network, which extracts multi-scale features through five convolutional stages and residual connections. The attention enhancement module is a CBAM convolutional block attention module, which assigns weights to the feature map through channel attention sub-modules and spatial attention sub-modules in sequence. The classification output module outputs the binary classification result and confidence score through global average pooling, two fully connected layers and a Sigmoid activation function.

[0027] In one possible implementation, the deep learning model is built around a "ResNet50 residual network + CBAM dual attention mechanism." The input layer receives a 128×128×3 three-dimensional feature matrix. The feature extraction module uses a ResNet50 network, which alleviates the gradient vanishing problem through five convolutional stages and residual connection structures, deeply mining the local texture, mid-level structure, and global semantic features of the echo. A CBAM attention module is embedded after the feature extraction module. First, a channel attention submodule performs global average pooling and global max pooling on the feature map. The results are input into a shared fully connected layer and activated by a Sigmoid function to generate channel weight vectors, strengthening the effective feature channel weights. Then, a spatial attention submodule performs average pooling and max pooling on the channel-weighted feature map. After concatenation, the resulting image is processed by a 1×1 convolutional layer and activated by a Sigmoid function to generate a spatial weight map, focusing on key regions. The classification output module converts the attention-enhanced feature map into a one-dimensional vector through global average pooling. It then outputs the confidence score in the range of 0 to 1 through two fully connected layers (mapped to 256-dimensional and 1-dimensional respectively) and the Sigmoid activation function. A confidence score ≥ 0.5 is considered a meteorological echo, otherwise it is considered a non-meteorological echo.

[0028] Step 104: Input the training set into the model, calculate the prediction bias using the binary cross-entropy loss function, and dynamically adjust the model parameters using the Adam adaptive momentum optimizer.

[0029] In addition, L2 regularization and early stopping strategies are introduced to suppress overfitting, and grid search is used to optimize key hyperparameters such as learning rate, kernel size, and batch size.

[0030] During the model training and optimization phases, the binary cross-entropy loss function is used to calculate the prediction bias, as shown in the formula: Where N is the number of samples, y i For real labels, p i To predict probabilities, the Adam adaptive momentum optimizer was used to dynamically adjust parameters. The initial learning rate was 0.001, and the weight decay coefficient was 0.0001 (L2 regularization). Every 50 rounds, the learning rate was decayed to 0.5 of its original value. A grid search method was used to optimize hyperparameters, traversing all combinations of learning rate (0.0001, 0.001, 0.01), kernel size (3×3, 5×5), and batch size (16, 32, 64). The optimal combination was finally determined to be a learning rate of 0.001, a kernel size of 3×3, and a batch size of 32.

[0031] Step 105: After standardizing the radar echo data to be identified according to the data acquisition and preprocessing steps, input it into the trained optimal model, and output the meteorological echo / non-meteorological echo classification results and the confidence level in the 0~1 interval.

[0032] The data acquisition steps include step 101, which involves acquiring raw weather radar echo data containing core parameters such as reflectivity factor, radial velocity, and spectral width. The data preprocessing steps include step 101, which involves performing format standardization conversion, combined noise removal using adaptive median filtering and wavelet thresholding, invalid value filling using neighborhood interpolation, and parameter normalization on the raw echo data to generate a three-dimensional echo feature matrix of uniform dimension.

[0033] During training, the accuracy of the validation set was monitored in real time. If there was no improvement for 10 consecutive rounds, an early stopping strategy was triggered. In this implementation, training was stopped at the 85th round, and the optimal model was saved. Model performance was validated on the test set, covering multiple scenarios including mountainous terrain echo interference, coastal wave echo interference, weak insect echoes in spring, and super-refractive echoes in winter. The results showed that the model could stably distinguish between the two types of echoes in all scenarios, with an inference latency consistently within 85ms, meeting the real-time processing requirement of ≤100ms. Furthermore, the recognition stability on C-band and S-band radar data was basically consistent, demonstrating excellent generalization ability.

[0034] After system deployment, the data acquisition module receives radar data streams from three observation stations in real time via TCP / IP protocol (transmission rate 10 frames / second). The data preprocessing module processes the data in parallel, the recognition and inference module outputs classification results and confidence levels, and the results display module displays echo parameter distribution maps, recognition heat maps (red lines indicate meteorological echoes, blue lines indicate non-meteorological echoes, and color depth corresponds to confidence levels) and classification statistics (including real-time recognition accuracy and sample category distribution ratios) in real time. The system supports exporting results as CSV files or image files. The system has been running continuously and stably for 720 hours without failure.

[0035] Compared with the prior art, the present invention achieves the following significant beneficial effects.

[0036] The reliability of recognition is significantly improved. The model that integrates ResNet50 and CBAM dual attention mechanism can automatically mine deep and complex features without human intervention. The recognition stability of difficult-to-distinguish echoes such as super-refractive and weak insects is greatly improved. It can accurately separate the two types of echoes in complex scenes and reduce the misjudgment of meteorological data. Through training with multi-source radar data, standardization processing, and data augmentation, the model can be adapted to different types of radar and complex geographical environments, and can maintain stable performance without adjusting parameters in unfamiliar scenarios. Through model structure optimization and system pipeline processing, the total time for single-frame data preprocessing + inference is ≤100ms, supporting real-time data stream processing of ≥10 frames / second, meeting the requirements for early warning timeliness; It achieves end-to-end automatic identification, and the operation process is simple, which can be mastered by non-technical personnel after simple training. The system covers the entire process and supports real-time processing and historical data backtracking. It can be directly connected to meteorological business systems. Its modular design facilitates the subsequent access of new parameters, network replacement, or addition of classification categories, adapting to the needs of technological development.

[0037] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0038] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware, or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. The above descriptions are merely optional embodiments of this application and are not intended to limit the application. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for identifying meteorological echoes and non-meteorological echoes from weather radar, characterized in that, include: Raw weather radar echo data containing core parameters such as reflectivity factor, radial velocity, and spectral width are acquired. The raw echo data undergoes format standardization, a combined noise removal process combining adaptive median filtering and wavelet thresholding, invalid value filling via neighborhood interpolation, and parameter normalization to generate a unified three-dimensional echo feature matrix. In the combined noise removal, the adaptive median filtering window size is 3×3~7×7, dynamically adjusted according to noise density. Wavelet thresholding uses the db4 wavelet basis function, and after three-level wavelet decomposition, soft thresholding is applied to high-frequency coefficients. The parameter normalization process maps parameters to the [0,1] interval, with the mapping formula being x... norm =(xx min ) / (x max -x min ), where x is the original parameter value, x min x max The global minimum and maximum values ​​of this parameter in historical data; the search range for invalid value filling by the neighborhood interpolation method is a 3×3~5×5 neighborhood area around the invalid value pixel. The filling is completed by calculating the weighted average of the effective pixels in the neighborhood. The weight decreases linearly with the distance of the pixel from the invalid value point. The three-dimensional echo feature matrix is ​​manually labeled, with the labeling category being either meteorological echo or non-meteorological echo. The labeling results are verified for consistency, and samples with a labeling consistency rate ≥98% are included in the effective sample set. The training set, validation set, and test set are divided in a 7:1:2 ratio. The training set samples are expanded using multi-dimensional data augmentation techniques, including random rotation, horizontal / vertical flipping, Gaussian noise addition, and local cropping. In the multi-dimensional data augmentation techniques, the random rotation angle is 0°, 90°, 180°, or 270°, the variance of the Gaussian noise ranges from 0.01 to 0.05, and the local cropping ratio ranges from 0.8 to 1.

0. After data augmentation, the number of training set samples is expanded by 3 to 5 times. A deep convolutional neural network model is constructed, which sequentially includes an input layer, a feature extraction module, an attention enhancement module, and a classification output module. The input layer is used to receive the three-dimensional echo feature matrix. The feature extraction module adopts a ResNet50 residual network and extracts multi-scale features through five convolutional stages and residual connections. The attention enhancement module is a CBAM convolutional block attention module, which assigns weights to the feature map through channel attention sub-modules and spatial attention sub-modules in sequence. The classification output module outputs the binary classification result and confidence score through global average pooling, two fully connected layers, and a sigmoid activation function. The training set is input into the model, and the prediction bias is calculated using the binary cross-entropy loss function. The model parameters are dynamically adjusted using the Adam adaptive momentum optimizer. L2 regularization and early stopping strategies are introduced to suppress overfitting. The key hyperparameters, such as learning rate, kernel size, and batch size, are optimized using a grid search method. The initial learning rate of the Adam adaptive momentum optimizer is 0.001, and the weight decay coefficient is 0.0001. A learning rate decay strategy is adopted, which reduces the learning rate to 0.5 every 50 rounds. The early stopping strategy is triggered when the accuracy of the validation set does not improve for 10 consecutive rounds. The hyperparameter optimization range of the grid search method is: learning rate 0.0001~0.01, kernel size 3×3 or 5×5, batch size 16~64, and training iterations 50~200 rounds. After standardizing the radar echo data to be identified according to the data acquisition and preprocessing steps, the data is input into the trained optimal model, and the meteorological echo / non-meteorological echo classification results and the confidence level in the 0~1 interval are output.

2. The method according to claim 1, characterized in that, The format standardization conversion transforms the raw data of IRIS and NetCDF formats into a unified H×W×3 three-dimensional matrix, where H and W are the height and width of the echo image, with values ​​ranging from 64 to 256, and the three channels correspond to the reflectivity factor, radial velocity, and spectral width parameters, respectively.

3. The method according to claim 1, characterized in that, In the combined noise removal, the window size of the adaptive median filter is 3×3~7×7, which is dynamically adjusted according to the noise density; Wavelet thresholding denoising uses the db4 wavelet basis function, and after 3-level wavelet decomposition, soft thresholding is applied to the high-frequency coefficients. The parameter normalization process involves mapping the parameters to the [0,1] interval, and the mapping formula is x. norm =(xx min ) / (x max -x min ), where x is the original parameter value, x min x max These are the global minimum and maximum values ​​of this parameter in historical data.

4. The method according to claim 1, characterized in that, The search range for invalid value filling in the neighborhood interpolation method is a 3×3~5×5 neighborhood area around the invalid value pixel. The filling is completed by calculating the weighted average of the effective pixels in the neighborhood. The weight decreases linearly with the distance of the pixel from the invalid value point.

5. The method according to claim 1, characterized in that, In the aforementioned multi-dimensional data augmentation technique, the random rotation angle is 0°, 90°, 180° or 270°, the variance of Gaussian noise is in the range of 0.01~0.05, the local cropping ratio is in the range of 0.8~1.0, and the number of training set samples is increased by 3~5 times after data augmentation.

6. The method according to claim 1, characterized in that, The channel attention submodule performs global average pooling and global max pooling on the feature map, inputs the two pooling results into a shared fully connected layer, and generates a channel weight vector through Sigmoid activation; the spatial attention submodule performs average pooling and max pooling on the channel-weighted feature map, concatenates the pooling results, and generates a spatial weight map through a convolutional layer and Sigmoid activation.

7. The method according to claim 1, characterized in that, The Adam adaptive momentum optimizer has an initial learning rate of 0.001 and a weight decay coefficient of 0.0001. It adopts a learning rate decay strategy that reduces the learning rate to 0.5 every 50 rounds. The early stopping strategy is triggered when the validation set accuracy does not improve for 10 consecutive rounds. The hyperparameter optimization range of the grid search method is: learning rate 0.0001~0.01, kernel size 3×3 or 5×5, batch size 16~64, and training iterations 50~200 rounds.

8. A weather radar meteorological echo and non-meteorological echo identification system, applicable to the weather radar meteorological echo and non-meteorological echo identification method according to any one of claims 1 to 7, characterized in that, The system includes: The data acquisition module is used to acquire raw echo data from weather radar containing core parameters such as reflectivity factor, radial velocity, and spectral width. It performs format standardization conversion, combined noise removal using adaptive median filtering and wavelet thresholding, invalid value filling via neighborhood interpolation, and parameter normalization on the raw echo data to generate a unified three-dimensional echo feature matrix. In the combined noise removal, the window size of the adaptive median filtering is 3×3~7×7, dynamically adjusted according to the noise density. Wavelet thresholding uses the db4 wavelet basis function, and after three-level wavelet decomposition, soft thresholding is applied to the high-frequency coefficients. The parameter normalization process maps the parameters to the [0,1] interval, with the mapping formula being x... norm =(xx min ) / (x max -x min ), where x is the original parameter value, x min x max The global minimum and maximum values ​​of this parameter in historical data; the search range for invalid value filling by the neighborhood interpolation method is a 3×3~5×5 neighborhood area around the invalid value pixel. The filling is completed by calculating the weighted average of the effective pixels in the neighborhood. The weight decreases linearly with the distance of the pixel from the invalid value point. The sample management module is used to manually annotate the three-dimensional echo feature matrix, with the annotation category being either meteorological echo or non-meteorological echo. The module performs consistency verification on the annotation results, and includes samples with an annotation consistency rate ≥98% into the effective sample set. The training set, validation set, and test set are divided in a 7:1:2 ratio. The training set samples are expanded using multi-dimensional data augmentation techniques, including random rotation, horizontal / vertical flipping, Gaussian noise addition, and local cropping. In these multi-dimensional data augmentation techniques, the random rotation angle is 0°, 90°, 180°, or 270°, the variance of the Gaussian noise ranges from 0.01 to 0.05, and the local cropping ratio ranges from 0.8 to 1.

0. After data augmentation, the number of training set samples is increased by 3 to 5 times. The model training module is used to construct a deep convolutional neural network model that sequentially includes an input layer, a feature extraction module, an attention enhancement module, and a classification output module. The input layer is used to receive a three-dimensional echo feature matrix. The feature extraction module uses a ResNet50 residual network and extracts multi-scale features through five convolutional stages and residual connections. The attention enhancement module is a CBAM convolutional block attention module, which assigns weights to the feature map through channel attention sub-modules and spatial attention sub-modules in sequence. The classification output module outputs the binary classification result and confidence score through global average pooling, two fully connected layers, and a Sigmoid activation function. The bias prediction module inputs the training set into the model, calculates the prediction bias using the binary cross-entropy loss function, dynamically adjusts the model parameters using the Adam adaptive momentum optimizer, introduces L2 regularization and an early stopping strategy to suppress overfitting, and optimizes key hyperparameters such as learning rate, kernel size, and batch size using a grid search method. The initial learning rate of the Adam adaptive momentum optimizer is 0.001, the weight decay coefficient is 0.0001, and a learning rate decay strategy of reducing the learning rate to 0.5 every 50 rounds is adopted. The early stopping strategy is triggered when the accuracy of the validation set does not improve for 10 consecutive rounds. The hyperparameter optimization range of the grid search method is: learning rate 0.0001~0.01, kernel size 3×3 or 5×5, batch size 16~64, and training iterations 50~200 rounds. The model output module is used to standardize the radar echo data to be identified according to the data acquisition and preprocessing steps, input it into the trained optimal model, and output the meteorological echo / non-meteorological echo classification results and the confidence level in the 0~1 interval.

Images