Weather radar gust front automatic identification system based on deep learning

Abstract

The invention relates to the technical field of automatic detection and early warning of weather radar gust front, and discloses a weather radar gust front automatic identification system based on deep learning, which comprises a radar data acquisition and preprocessing module used for acquiring radar echo data from weather radar equipment, carrying out format conversion on the data, and sending the converted data to a server; the radar reflectivity is converted from a polar coordinate format to a Cartesian coordinate format, and meanwhile, normalization processing of the radar reflectivity value is carried out; and the feature extraction and region candidate generation module is used for performing feature extraction on the radar image by adopting a deep convolutional neural network and generating a target region candidate box through a region candidate network, and the region candidate network comprises a classification branch and a regression branch. According to the method, rapid and accurate gust front identification is realized through a deep learning target detection and mask segmentation technology, the stability and robustness of detection are remarkably improved, false alarms and missing alarms are effectively reduced through dynamic alarm threshold calculation and trend analysis, and the accuracy and timeliness of alarm are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic detection and early warning of gust fronts by weather radars, and specifically to an automatic recognition system for weather radar gust fronts based on deep learning. Background Art

[0002] In the field of detecting weather radar gust fronts, the existing technologies mainly rely on two methods for analysis: one is manual recognition by meteorological experts based on experience, and the other is to extract gust front features using traditional image processing algorithms and perform detection. Although the manual recognition method has a certain degree of reliability in specific situations, due to its high dependence on the subjective judgment of experts, the stability of the recognition results is difficult to guarantee, and the analysis process is time-consuming and laborious, unable to meet the requirements of real-time early warning. On the other hand, the methods based on traditional image processing technologies use means such as edge detection, threshold segmentation, or morphological analysis to identify the gust front area, and rely on fixed alarm thresholds to trigger early warnings. However, when faced with the complexity and dynamic changes of radar echo data, such methods often appear powerless, resulting in low detection accuracy and alarm reliability.

[0003] In terms of the specific structure of the existing technologies, they usually include modules such as radar data acquisition, feature extraction, fixed threshold alarm, and static result analysis. However, these solutions have many deficiencies and are difficult to meet the actual application requirements. First of all, the manual recognition method leads to low detection efficiency and is easily affected by human factors, unable to guarantee the consistency of detection results. Secondly, the traditional image processing methods have poor anti-interference ability. When there is noise in the radar echo signal, it is easy to have false detections or missed detections, affecting the accuracy of detection. In addition, the existing alarm mechanisms rely on fixed thresholds and are difficult to adapt to different weather environments, resulting in insufficient alarm sensitivity during strong storms and frequent false alarms during normal weather fluctuations, reducing the practicality of the system. At the same time, there is a lack of trend analysis means and it is impossible to predict the movement trajectory of the gust front, resulting in delayed early warnings and affecting the effect of disaster prevention and mitigation. Finally, the output method of the alarm signal is single and does not provide graded alarms according to different danger levels, making it difficult for users to quickly make reasonable response measures. These problems limit the application scope of the existing technologies, and there is an urgent need for a more intelligent and stable detection and alarm scheme to improve the reliability and practicality of the system. Summary of the Invention

[0004] Aiming at the deficiencies of the existing technologies, the present invention provides an automatic recognition system for weather radar gust fronts based on deep learning, which solves the problems of low efficiency of manual recognition, poor anti-interference of image processing methods, false alarms and missed detections caused by fixed alarm mechanisms, and lack of dynamic trend analysis in the existing technologies.

[0005] To achieve the above objectives, the present invention is realized through the following technical solutions: An automatic recognition system for weather radar gust fronts based on deep learning, including:

[0006] A radar data acquisition and preprocessing module, which is used to collect radar echo data from weather radar equipment, perform format conversion on the data, convert it from polar coordinate format to Cartesian coordinate format, and simultaneously perform normalization processing on radar reflectivity values;

[0007] A feature extraction and region candidate generation module, which uses a deep convolutional neural network to extract features from radar images and generates target region candidate boxes through a region candidate network. The region candidate network includes a classification branch and a regression branch;

[0008] A target detection and mask segmentation module, which is used to perform target detection on candidate regions, classify and generate a pixel-level mask image of the target. The module includes a fully convolutional network for generating target masks;

[0009] A model training and optimization module, which uses an optimization algorithm based on gradient descent to train a deep learning model and optimizes network parameters through a loss function. The loss function includes classification loss, bounding box regression loss, and mask segmentation loss;

[0010] A detection and alarm output module, which is used to perform real-time detection of gust fronts through real-time input radar data and a trained model, and trigger an alarm mechanism to send warning messages to relevant meteorological personnel.

[0011] Preferably, the radar data acquisition and preprocessing module converts radar data from polar coordinate format to Cartesian coordinate format by performing format conversion on radar echo data, and performs normalization processing on reflectivity data to convert it into a value within a standard range.

[0012] Preferably, the feature extraction and region candidate generation module includes a deep convolutional neural network, and the network is a ResNet network, which is used to extract high-level features in radar images, and generates target region candidate boxes through a region candidate network. The network determines whether the candidate region contains a gust front target through the classification branch, and optimizes the boundary of the candidate box through the regression branch.

[0013] Preferably, the target detection and mask segmentation module adopts a MaskR-CNN architecture, which includes a feature extraction network, a region candidate network, RoIAlign, a fully connected layer, and a fully convolutional layer, and is used to generate a pixel-level mask image of the target region and classify the target.

[0014] Preferably, the model training and optimization module uses an optimization algorithm to train a deep learning model, and the optimization algorithm is used to optimize the loss function according to training data. The loss function includes classification loss, bounding box regression loss, and mask segmentation loss.

[0015] Preferably, the detection and alarm output module includes a real-time data acquisition module that receives radar data in real time, preprocesses it, and then inputs it into a trained deep learning model for gust front detection, and sends an alarm to relevant personnel by means of SMS, email, or APP push.

[0016] Preferably, the deep convolutional neural network adopts a residual neural network, and the network includes multiple residual blocks, and each residual block includes a layer of convolution operation and a layer of activation function.

[0017] Preferably, the radar data acquisition and preprocessing module further includes a data augmentation module, and the module enhances the training dataset by means of rotation, flipping, and adding noise.

[0018] Preferably, the system outputs the detection results in real time through a graphical user interface, and the results include the detection position, bounding box, and mask image of the gust front, as well as relevant meteorological information.

[0019] Preferably, the object detection and mask segmentation module generates an object mask through a fully convolutional network.

[0020] The present invention provides an automatic gust front recognition system for weather radar based on deep learning, which has the following beneficial effects:

[0021] 1. The present invention adopts object detection and mask segmentation technology based on deep learning to automatically identify the gust front area in radar echo data, achieving the technical effects of fast and accurate identification of the gust front area. Compared with the existing solutions that rely on manual identification or traditional image processing algorithms, it solves the problems of low recognition accuracy, slow efficiency, and susceptibility to noise interference, and significantly improves the stability and robustness of detection.

[0022] 2. The present invention introduces a dynamic alarm threshold calculation mechanism, and adaptively adjusts the alarm trigger condition by using the mean and standard deviation of real-time data, achieving the technical effects of reducing false alarms and avoiding missed alarms under complex weather conditions. Compared with the existing alarm mechanism with a fixed threshold, it solves the problem of insufficient alarm sensitivity in sudden weather changes, and significantly improves the reliability of the system.

[0023] 3. The present invention adopts time series analysis and Kalman filtering algorithms to predict and analyze the movement trend of the gust front, achieving the technical effects of improving the timeliness and accuracy of alarms. Compared with the traditional static analysis method based on single-frame images, it solves the problem of lagging response to rapidly changing weather phenomena, and significantly enhances the early warning ability of the system.

[0024] 4. The present invention introduces a multi-level alarm output mechanism. By combining the characteristics of radar echo data and multi-source data judgment, it realizes the hierarchical output of alarm information, achieving the technical effect of flexibly adjusting the alarm intensity according to the degree of danger. Compared with the traditional single alarm mode, it solves the problems of lack of pertinence of alarm information and difficulty for users to respond quickly, significantly improving the practicality of the system and user friendliness. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is the system architecture diagram of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0026] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0027] Please refer to the attached Figure 1 , the embodiment of the present invention provides an automatic recognition system for gust fronts of weather radar based on deep learning, including:

[0028] A radar data acquisition and preprocessing module, which is used to collect radar echo data from weather radar equipment, perform format conversion on the data, convert it from polar coordinate format to Cartesian coordinate format, and at the same time perform normalization processing on the radar reflectivity value;

[0029] In the automatic recognition system for gust fronts of weather radar based on deep learning of the present invention, the acquisition and preprocessing of radar data are the basis of the entire detection process. This module is responsible for obtaining echo data from the weather radar system, and performing format conversion, normalization, denoising, and data enhancement on it to meet the requirements of the deep learning model for input data.

[0030] Generally, the weather radar system adopts a pulse Doppler system, emits electromagnetic waves of a specific frequency, and receives the echo signals in the target area. Radar data is usually stored in polar coordinates, and each data point is determined by the distance r and the azimuth angle θ together, reflecting the radar reflectivity intensity in the target area. Since the storage format of radar data does not match the input format of the deep learning model, coordinate conversion must be performed to make it conform to the standard Cartesian coordinate format. In addition, radar signals will be affected by factors such as meteorological noise and ground clutter interference during the propagation process, resulting in a decline in data quality. Therefore, it is necessary to perform normalization processing and denoising operations on the data to improve data quality and model adaptability.

[0031] Data Format Conversion

[0032] In this embodiment, the radar data format conversion adopts the bilinear interpolation method to convert the polar coordinate data into the Cartesian coordinate format.

[0033] Specifically, the data point P(r,θ) in the polar coordinate system needs to be mapped to the point P(x,y) in the Cartesian coordinate system, and the conversion formula is as follows:

[0034] x = rcosθ;

[0035] y = rsinθ;

[0036] Among them, x and y are the horizontal and vertical coordinates in the Cartesian coordinate system;

[0037] r is the radial distance in the radar data;

[0038] θ is the azimuth angle of the radar data, and the unit is radian.

[0039] In a possible implementation manner, in order to improve the accuracy of data conversion, the bilinear interpolation method is used to calculate the echo value I(x,y) on the Cartesian grid:

[0040] I(x,y) = (1-α)(1-β)I(i,j) + α(1-β)I(i+1,j) + (1-α)βI(i,j+1) + αβI(i+1,j+1);

[0041] Among them,

[0042] I(x,y) is the echo intensity value on the transformed Cartesian coordinate grid;

[0043] I(i,j), I(i+1,j), I(i,j+1), I(i+1,j+1) are the echo values of the four adjacent data points in the polar coordinate grid;

[0044] α and β are the interpolation weights, and the calculation methods are as follows:

[0045]

[0046]

[0047] Among them,

[0048] Take the integer part of x;

[0049] Take the integer part of y.

[0050] Normalization processing

[0051] In some embodiments, in order to make the data of different radar devices consistent and improve the stability of model training, the radar data can be normalized.

[0052] Specifically, in this embodiment, Z-score normalization is adopted, and the calculation formula is as follows:

[0053]

[0054] Wherein,

[0055] I norm (x, y) is the radar echo value after normalization;

[0056] I(x, y) is the original echo value before normalization;

[0057] μ is the mean value of radar data, and the calculation method is as follows:

[0058]

[0059] σ is the standard deviation of radar data, and the calculation method is as follows:

[0060]

[0061] N is the total number of radar echo values in the dataset.

[0062] In another possible implementation, min-max normalization can be selected, and its calculation formula is as follows:

[0063]

[0064] Wherein,

[0065] I min is the minimum radar echo value in the dataset;

[0066] I max is the maximum radar echo value in the dataset.

[0067] Denoising processing

[0068] In this embodiment, in order to reduce the noise influence in radar data and improve data quality, a method combining median filtering and bilateral filtering is adopted.

[0069] Specifically, the calculation method of median filtering is as follows:

[0070] I filtered (x, y) = median(I(i, j)), (i, j) ∈ window(x, y);

[0071] Wherein,

[0072] I filtered (x, y) is the radar data value after median filtering;

[0073] window(x,y) is the set of data points within a k×k window centered at (x,y).

[0074] As an option, bilateral filtering can be further employed:

[0075] I filtered (x,y) = ∑ i,j W s (i,j)W r (i,j)I(i,j);

[0076] where,

[0077] W s (i,j) is the spatial weight, calculated as follows:

[0078]

[0079] W r (i,j) is the intensity weight, calculated as follows:

[0080]

[0081] σ s is the spatial smoothing parameter;

[0082] σ r is the intensity smoothing parameter.

[0083] Data augmentation

[0084] In some embodiments, to improve the generalization ability of the model, data augmentation techniques such as rotation transformation, mirror flipping, Gaussian noise addition, etc. can be employed.

[0085] In one possible implementation, the calculation method of the rotation transformation is as follows:

[0086] x′ = xcosθ - ysinθ;

[0087] y′ = xsinθ + ycosθ;

[0088] where,

[0089] x′, y′ are the rotated coordinates;

[0090] θ is the rotation angle, in radians. In addition, Gaussian noise can be added:

[0091]

[0092] where,

[0093] is the noise that follows a normal distribution with a mean of 0 and a standard deviation of σ n ​

[0094] In summary, in this embodiment, through format conversion, normalization, denoising, and data augmentation, it is ensured that the radar data meets the input requirements of the deep learning model, improving the data quality and the stability of detection.

[0095] The feature extraction and region candidate generation module uses a deep convolutional neural network to extract features from the radar image and generates target region candidate boxes through a region candidate network, where the region candidate network includes a classification branch and a regression branch;

[0096] In the automatic weather radar gust front recognition system based on deep learning of the present invention, the feature extraction and region candidate generation module, as a key component, is responsible for extracting effective features from the preprocessed radar data and generating potential gust front region candidates. This module is closely connected with the aforementioned radar data acquisition and preprocessing module, uses the preprocessed and standardized radar echo data for feature extraction, and then generates region candidate boxes, providing a basis for subsequent target detection and classification. During the feature extraction process, advanced deep learning methods such as convolutional neural networks (CNNs) are used to mine representative spatial features from complex radar echo data, ensuring that the generated region candidate boxes can accurately cover possible gust front regions.

[0097] Generally, after being processed by the preprocessing module, the radar echo data has been converted into a format suitable for input into the deep learning model and has undergone operations such as normalization, denoising, and enhancement. These processes make the radar echo data more stable and have better feature expression ability. The main objective of the feature extraction and region candidate generation module is to further extract features related to gust front recognition based on this data and generate candidate regions based on these features. Feature extraction not only involves traditional image processing methods but also leverages the advantages of deep learning networks, being able to automatically learn complex spatial patterns and features in the data.

[0098] In this embodiment, the feature extraction and region candidate generation module mainly includes the following steps: feature extraction, feature map generation, region candidate box generation, and candidate box screening.

[0099] Feature Extraction

[0100] As an option, the feature extraction part uses a convolutional neural network (CNN), which can effectively extract local spatial features from the radar data. In a specific implementation, the CNN contains multiple convolutional layers and pooling layers, and local feature extraction is performed on the radar image through a sliding convolutional kernel. The calculation formula for the convolution operation is as follows:

[0101] I conv (x,y) = ∑ i,j K(i,j)I(x + i,y + j);

[0102] Among them,

[0103] I conv (x, y) is the value of the feature map after the convolution operation;

[0104] I(x + i, y + j) is the value of the input radar data;

[0105] K(i, j) is the weight value of the convolution kernel;

[0106] x, y are the current coordinates of the convolution operation;

[0107] i, j are the indices of the convolution kernel.

[0108] Specifically, the convolutional layer performs weighted summation on local regions of the radar data to extract low-level features in the image, such as edges, textures, etc. Subsequently, the pooling layer downsamples the feature map, reduces the data dimension, improves the computational efficiency of the network, and has a certain translational invariance. The pooling operation usually uses max pooling or average pooling. The calculation formula for the max pooling operation is:

[0109] I pool (x, y) = max(I(x + i, y + j)), (i, j) ∈ pooling window;

[0110] Among them,

[0111] I pool (x, y) is the value of the feature map after pooling;

[0112] (i, j) are the data points in the pooling window.

[0113] Through multiple convolution and pooling operations, the CNN can extract higher-level features from the radar data, such as potential features of targets like weather patterns, gust fronts, etc.

[0114] Region proposal generation

[0115] After feature extraction, the next step is to generate candidate regions that may contain gust fronts. To generate high-quality region proposals, a sliding window method based on the feature map is used to generate candidate boxes of different sizes. Specifically, a window of a fixed size is slid in the feature map, and the feature response value within the window is calculated. Assuming the size of the candidate box is w × h, the feature response value R(x, y) at the window position (x, y) can be expressed as:

[0116]

[0117] Among them,

[0118] R(x, y) is the response value of the candidate box at position (x, y);

[0119] W(i, j) is the weight coefficient related to the window size;

[0120] I conv (x + i, y + j) is the value of the feature map after convolution.

[0121] By sliding the window over the entire feature map, a series of candidate regions are generated. Each candidate region contains a local region of the radar data, and this region may contain gust front targets.

[0122] Region candidate box screening

[0123] Since the generated region candidate boxes may contain a large number of irrelevant regions, it is necessary to further screen out the regions most likely to contain gust fronts. The screening process usually includes the non - maximum suppression (NMS) algorithm, which can remove candidate boxes with a high degree of overlap. Specifically, the NMS algorithm sorts the candidate boxes according to their confidence levels and gradually removes the boxes with a high degree of overlap with the current box and a lower confidence level. The overlap degree (intersection over union, IoU) of the candidate boxes is set as:

[0124]

[0125] Where

[0126] A and B are two candidate boxes;

[0127] |A ∩ B| is the area of the intersection region of candidate boxes A and B;

[0128] |A ∪ B| is the area of the union region of candidate boxes A and B.

[0129] If IoU(A, B) is greater than the set threshold, it is considered that the overlap degree of these two boxes is too high. Remove the boxes with a high degree of overlap and retain the box with a higher confidence level.

[0130] Feature map generation and subsequent processing

[0131] In some embodiments, after the feature map output by the deep convolutional network for feature extraction undergoes region candidate box generation and screening, it will be input into the subsequent object detection network for object recognition and classification. In a possible implementation, the method of region - based convolutional neural network (R - CNN) is combined to classify the candidate boxes and further refine the position of the bounding boxes. Specifically, the region - based convolutional neural network performs convolution on the features within each candidate box again to extract more fine - grained features for object classification.

[0132] In summary, the feature extraction and region candidate generation module of this embodiment extracts spatial features in radar data through a convolutional neural network, and generates region candidate frames that may contain gust fronts based on these features. Through subsequent screening and processing, the gust front region in the radar data can be effectively identified, and reliable data support can be provided for subsequent target detection and classification.

[0133] The target detection and mask segmentation module is used to detect the target in the candidate area, classify it and generate the pixel-level mask image of the target. The module includes a fully convolutional network to generate the target mask.

[0134] In the weather radar gust front automatic identification system based on deep learning of the present invention, the target detection and mask segmentation module is used to accurately identify the candidate area provided by the area candidate generation module, and perform pixel-level segmentation on the specific boundary of the gust front. This module is closely connected with the aforementioned feature extraction and area candidate generation module, with the deep learning model as the core, to realize the detection of candidate areas and pixel-level mask generation. The detection part mainly locates the possible area of ​​the gust front through the target detection network, while the mask segmentation part further refines the boundary, making the detection result more accurate and accurately describing the shape of the gust front.

[0135] Generally speaking, the echo data of weather radar has strong spatial characteristics, and its echo pattern can reflect information such as storm structure and gust front shape. In the target detection stage, the model needs to be able to identify key areas in radar images and classify different types of meteorological phenomena. Since gust fronts usually have irregular boundaries, traditional rectangular bounding boxes are difficult to accurately describe their shapes. Therefore, the mask segmentation part uses semantic segmentation or instance segmentation methods to obtain more accurate regional contours.

[0136] In this embodiment, the target detection and mask segmentation module consists of a target detection network, a mask segmentation network and a post-processing part.

[0137] Object Detection Network

[0138] As an option, the target detection network uses a region proposal method (such as FasterR-CNN or YOLO) to identify gust fronts in radar data. In one possible implementation, a FasterR-CNN model is used, which consists of a backbone network (Backbone), a region proposal network (RPN), and a target classification and bounding box regression module.

[0139] Specifically, the input radar feature map I feat (x, y) extracts features through the backbone network, and the feature map is input to RPN to generate multiple candidate regions. The goal of RPN is to learn a set of region proposal boxes (x c ,y c, w, h), and its calculation method is as follows:

[0140]

[0141] Among them,

[0142] (x, y) is the upper left corner coordinate of the candidate box;

[0143] (w, h) is the width and height of the candidate box;

[0144] (x c , y c ) is the center coordinate of the transformed candidate box;

[0145] (w′, h′) is the size of the candidate box after regression optimization;

[0146] t w , t h is the bounding box regression parameter.

[0147] After the candidate boxes are filtered by non-maximum suppression (NMS), they are input into the classification network for target class prediction, and the candidate boxes are optimized to obtain a more accurate position.

[0148] In another possible implementation, the YOLO (You Only Look Once) model directly divides the grid on the feature map, and each grid predicts a fixed number of bounding boxes and their corresponding classes. Specifically, the prediction formula of YOLO is as follows:

[0149]

[0150]

[0151]

[0152] Among them,

[0153] is the confidence of the target existence;

[0154] σ(·) is the Sigmoid function;

[0155] (c x , c y ) is the offset of the grid cell;

[0156] f c , f x , f y , f w , f h is the parameter predicted by the network.

[0157] Mask segmentation network

[0158] Based on object detection, the mask segmentation part further generates pixel-level segmentation results to accurately depict the boundary of the gust front. As an option, the MaskR-CNN network can be adopted, which adds a mask branch on the basis of Faster R-CNN.

[0159] Specifically, the input feature map is first aligned through ROIAlign to ensure that the extracted features can maintain accuracy. Then, the mask branch uses a fully convolutional network (FCN) to output a binary mask that matches the candidate box:

[0160] M(x,y) = σ(W m *I roi (x,y));

[0161] Where,

[0162] M(x,y) is the predicted mask value (0 or 1);

[0163] I roi (x,y) is the input feature map of the ROI region;

[0164] W m is the convolutional weight of the mask branch;

[0165] σ(·) is the Sigmoid activation function.

[0166] As an option, another possible implementation is to use U-Net or DeepLabV3+ for semantic segmentation. These networks can perform global prediction on the entire radar image and generate the mask region of the gust front. U-Net adopts an encoder-decoder structure. The encoding part is used to extract high-level features, and the decoding part gradually restores the resolution. Its mask prediction formula is:

[0167] M(x,y) = softmax(W d *F decode (x,y));

[0168] Where,

[0169] F decode (x,y) is the output feature in the decoding stage;

[0170] W d is the weight parameter of the decoder;

[0171] function, for pixel-level classification.

[0172] Post-processing optimization

[0173] After mask segmentation is completed, the post-processing part optimizes the detection and segmentation results. Generally, it is necessary to further smooth the boundaries and eliminate mis-detected regions. As an option, morphological operations (such as closing) can be used to fill small holes in the mask to make the boundaries more coherent.

[0174] In a possible implementation, an optimization method based on boundary comparison can be used to calculate the boundary gradient of the mask region:

[0175]

[0176] where

[0177] G(x,y) is the gradient intensity of the mask boundary;

[0178] are the derivatives of the mask in the x and y directions.

[0179] If the gradient is large, it is considered that there is an obvious boundary in this region, and the contour of the mask can be further optimized through an edge filter.

[0180] In summary, the object detection and mask segmentation module of this embodiment identifies the potential region of the gust front through the object detection network, and further uses the mask segmentation method to accurately draw the boundary of the gust front. Through post-processing optimization, the detection result is made smoother, ensuring that the system can efficiently and accurately identify the gust front region in the radar image.

[0181] The model training and optimization module uses an optimization algorithm based on gradient descent to train the deep learning model, and optimizes the network parameters through a loss function, where the loss function includes classification loss, bounding box regression loss, and mask segmentation loss;

[0182] In the automatic identification system of the weather radar gust front based on deep learning, the model training and optimization module is crucial. This module models the radar data through deep learning methods, ensuring that the system can accurately identify the gust front region, reduce mis-detection and missed detection, and improve stability and generalization ability. Generally, model training includes multiple links such as data preprocessing, network structure design, optimization strategy adjustment, and model evaluation, and is closely connected with the object detection and alarm module.

[0183] Data Preprocessing

[0184] In this embodiment, the radar echo data usually has a large dynamic range and noise interference, so data preprocessing is the basis of model training. Generally, logarithmic normalization can be performed on the data to reduce the influence of extreme values and improve the learning stability of the model.

[0185] In addition, Gaussian filtering can also be used for data smoothing to reduce noise interference and improve the clarity of the target region.

[0186] To improve the generalization ability of the model, data augmentation is one of the key measures. Methods such as random rotation, scale scaling, and adding Gaussian noise can be used to simulate the changes in radar data under different weather conditions, enabling the model to adapt to complex environments.

[0187] Network architecture design

[0188] The choice of network architecture directly affects the detection accuracy and computational efficiency of the model. Generally, improved YOLO or Mask R-CNN models can be used for object detection and mask segmentation. Common backbone network choices include ResNet, EfficientNet, or Swin Transformer to balance detection accuracy and computational complexity.

[0189] In some embodiments, a dual-branch network architecture can be adopted:

[0190] The first branch is used for object detection to identify the gust front bounding boxes in the radar data;

[0191] The second branch is used for mask segmentation to extract a more refined gust front region and improve the detection accuracy.

[0192] In addition, to improve the feature extraction ability of the model, attention mechanisms (such as SE module or CBAM) can be combined to enhance the model's attention to key regions and improve the detection performance.

[0193] Training strategy optimization

[0194] The optimization of the training strategy directly affects the convergence speed and detection accuracy of the model. Generally, the loss function needs to consider both object detection and mask segmentation simultaneously, and a joint loss function can be adopted:

[0195] L = λ1L cls + λ2L box + λ3L mask ;

[0196] Where,

[0197] L cls represents the object classification loss,

[0198] L box represents the bounding box regression loss,

[0199] L mask represents the mask segmentation loss, and the weight coefficients of each part can be adjusted according to the task requirements.

[0200] In addition, a learning rate scheduling strategy such as cosine annealing learning rate adjustment can be adopted to optimize the training process and improve the stability and generalization ability of the model.

[0201] In a possible implementation, a Mixed Precision Training strategy can be used to reduce the video memory occupancy, improve the training speed, and maintain the detection accuracy at the same time.

[0202] Model Evaluation and Adjustment

[0203] Model evaluation is a crucial step to ensure the stability of the trained model in practical applications. Generally, the following metrics can be used for evaluation:

[0204] Mean Average Precision (mAP): Measures the overall accuracy of the detection model;

[0205] Intersection over Union (IoU): Measures the overlap degree between the predicted bounding box and the true target;

[0206] Recall: Measures the ability of the model to capture the target.

[0207] In addition, to ensure the time generalization ability of the model, a sliding window test strategy can be adopted to evaluate the detection effect of radar data in different time periods.

[0208] To improve the adaptability of the model under different weather conditions, an adaptive data augmentation technique can be combined and tested in a low signal-to-noise ratio environment to verify its stability. Meanwhile, in a new radar data environment, transfer learning technology can be used for fine-tuning to improve the generalization ability of the model under different regions and weather conditions.

[0209] The model training and optimization module in this embodiment ensures the efficiency and stability of the weather radar gust front automatic recognition system through fine data preprocessing, optimized network structure, reasonable training strategy, and comprehensive model evaluation, providing reliable technical support for target detection and alarm output.

[0210] The detection and alarm output module is used to perform real-time detection of the gust front through the real-time input radar data and the trained model, and trigger the alarm mechanism to send warning information to relevant meteorological personnel;

[0211] In the weather radar gust front automatic recognition system, the detection and alarm output module is the final decision-making link, responsible for parsing the recognition results output by the model training and optimization module and generating corresponding alarm information. Generally, this module needs to combine the results of object detection and mask segmentation, evaluate the detected gust front area, and trigger a warning signal according to the set alarm threshold. This module is closely connected to the object detection module, and at the same time, it is necessary to ensure the real-time and accuracy of the alarm output to meet the actual application requirements.

[0212] In this embodiment, the detection and alarm output module mainly includes multiple parts such as data parsing, alarm threshold setting, alarm rule matching, and alarm signal output.

[0213] Generally, data parsing is the basic step of the detection and alarm output module. This module needs to receive the output of the target detection module, including the detected gust front area, bounding box coordinates, confidence score, and mask segmentation result. In a possible implementation, the overall influence range of the gust front can be evaluated by calculating the confidence-weighted area of different target regions:

[0214]

[0215] Among them,

[0216] A eff is the effective area weighted by confidence;

[0217] S i is the area of the i-th detection region;

[0218] C i is the confidence score of this region;

[0219] N is the number of detected target regions.

[0220] As an option, the detected regions can be further screened to eliminate misdetected targets with small area and low confidence, so as to improve the reliability of the alarm. Generally, an area threshold S min and a confidence threshold C min can be set, and only the targets that meet the following conditions are retained:

[0221] S i ≥S min ,C i ≥C min ;

[0222] Specifically, alarm threshold setting is the key link of the detection and alarm output module, which directly affects the alarm accuracy of the system. In some embodiments, fixed alarm thresholds can be set according to historical weather data and expert experience. For example, if the detected effective area A eff exceeds the set alarm threshold A th , an alarm signal is triggered:

[0223] A eff ≥A th ;

[0224] Among them,

[0225] A th is the set alarm trigger threshold.

[0226] In another possible implementation, an adaptive threshold setting method can be adopted. Generally, the alarm threshold can be dynamically adjusted according to the real-time statistical characteristics of weather radar data. For example, the sliding window method can be used to calculate the average effective area within the past T time steps:

[0227]

[0228] A avg is the average effective area within the past T time steps (unit: km 2 );

[0229] A eff (t) is the effective area at time step t (unit: km 2 );

[0230] T is the set size of the time window (unit: minutes or hours).

[0231] Then, the alarm threshold is adjusted according to the standard deviation:

[0232] A th = A avg + k·σ;

[0233] Among them,

[0234] A avg is the average effective area of the past T time steps;

[0235] σ is the standard deviation of the effective area;

[0236] k is the adjustment coefficient.

[0237] Alarm rule matching is the core part to ensure the rationality of alarm output. Generally, in addition to the alarm threshold judgment based on the effective area, other meteorological parameters such as the change trends of wind speed and wind direction can be combined to construct a comprehensive alarm strategy. For example, in some embodiments, if the detected gust front area shows an expanding trend and the wind speed change exceeds the set threshold V th , then a high-level alarm is preferentially triggered:

[0238] ΔV > V th ;

[0239] Among them,

[0240] is the change rate of the detection area area over time;

[0241] ΔV is the wind speed change amount in recent frames.

[0242] As an option, geographical location information can be further combined to give priority alarms for high-risk areas. For example, detection results near airports, seaports or densely populated areas can be given higher alarm weights to improve the practicality of the early warning system.

[0243] The output of the alarm signal is the final link of the detection and alarm output module. Generally, this module needs to transmit the alarm information to an external system, such as a meteorological monitoring center or a disaster early warning platform. In one possible implementation, a multi-level alarm mechanism can be adopted to divide alarm signals of different levels:

[0244] Low-level alarm: When the detected gust front area is small but may affect the local meteorological conditions, a prompt message is output;

[0245] Medium-level alarm: When the detected gust front area is large and the confidence level is high, a warning signal is triggered;

[0246] High-level alarm: When the detection result indicates that the gust front may have a significant impact on key areas (such as airports, seaports, etc.), an emergency alarm is immediately triggered.

[0247] In some embodiments, visualization technology can be further combined to superimpose the detected gust front area on the radar image and generate a dynamic early warning map for further analysis by meteorological experts.

[0248] In another possible implementation, wireless communication or cloud computing technology can be adopted to push the alarm signal to user terminals in real time, such as mobile applications or professional monitoring systems, to improve the dissemination efficiency of early warning information.

[0249] In summary, the detection and alarm output module in this embodiment ensures the real-time performance and accuracy of the weather radar gust front automatic recognition system through efficient data parsing, reasonable alarm threshold setting, intelligent alarm rule matching and multi-level alarm signal output, providing a solid technical support for meteorological early warning.

[0250] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made in these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. Weather radar gust front automatic identification system based on deep learning, characterized by: include: Radar data acquisition and preprocessing module, used to collect radar echo data from weather radar equipment, convert the data format from polar coordinate format to Cartesian coordinate format, and normalize the radar reflectivity value; The feature extraction and region candidate generation module uses a deep convolutional neural network to extract features from radar images and generates target region candidate frames through a region candidate network, which includes a classification branch and a regression branch. The target detection and mask segmentation module is used to detect the target in the candidate area, classify it and generate a pixel-level mask image of the target. The module includes a fully convolutional network for generating the target mask. A model training and optimization module, which trains the deep learning model using a gradient descent-based optimization algorithm and optimizes network parameters through loss functions, including classification loss, bounding box regression loss, and mask segmentation loss; The detection and alarm output module is used to detect gust fronts in real time through real-time input radar data and trained models, and trigger the alarm mechanism to send warning information to relevant meteorological personnel.

2. The weather radar gust front automatic identification system based on deep learning according to claim 1 is characterized in that: The radar data acquisition and preprocessing module converts the radar data from polar coordinate format to Cartesian coordinate format by performing format conversion on the radar echo data, and normalizes the reflectivity data to convert it into a value within a standard range.

3. The weather radar gust front automatic identification system based on deep learning according to claim 1 is characterized in that: The feature extraction and region candidate generation module includes a deep convolutional neural network, which is a ResNet network, which is used to extract high-level features in radar images and generate target region candidate frames through the region candidate network. The network determines whether the candidate region contains a gust front target through a classification branch, and optimizes the boundary of the candidate frame through a regression branch.

4. The weather radar gust front automatic identification system based on deep learning according to claim 1 is characterized in that: The target detection and mask segmentation module adopts the MaskR-CNN architecture, which includes a feature extraction network, a region candidate network, RoIAlign, a fully connected layer and a fully convolutional layer, and is used to generate a pixel-level mask image of the target area and classify the target.

5. The weather radar gust front automatic identification system based on deep learning according to claim 1 is characterized in that: The model training and optimization module uses an optimization algorithm to train the deep learning model, and the optimization algorithm is used to optimize the loss function according to the training data, and the loss function includes classification loss, bounding box regression loss and mask segmentation loss.

6. The weather radar gust front automatic identification system based on deep learning according to claim 1 is characterized in that: The detection and alarm output module includes a real-time data acquisition module, which receives radar data in real time, inputs the data into a trained deep learning model for gust front detection after preprocessing, and sends an alarm to relevant personnel via SMS, email or APP push.

7. The weather radar gust front automatic identification system based on deep learning according to claim 1 is characterized in that: The deep convolutional neural network adopts a residual neural network, and the network includes multiple residual blocks, each of which includes a layer of convolution operation and a layer of activation function.

8. The weather radar gust front automatic identification system based on deep learning according to claim 1 is characterized in that: The radar data acquisition and preprocessing module further includes a data enhancement module, which enhances the training data set by rotating, flipping, and adding noise.

9. The weather radar gust front automatic identification system based on deep learning according to claim 1 is characterized in that: The system outputs detection results in real time through a graphical user interface, the results including the detection position of the gust front, a bounding box and a mask image, and related meteorological information.

10. The weather radar gust front automatic identification system based on deep learning according to claim 1, characterized in that: The object detection and mask segmentation module generates object masks through a fully convolutional network.

Images