Intelligent gale early warning system and method based on radial velocity characteristics and storage medium
By using ring stitching of radar data and improved U-Net network segmentation technology, the problems of discontinuity in radar data boundaries and insufficient feature recognition capabilities have been solved, enabling accurate identification and automated early warning of mid-level radial convergence features, and improving the accuracy and timeliness of gale warnings.
Patent Information
- Application Number
- CN202610504516.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-16
- Publication Date
- 2026-07-10
AI Technical Summary
Existing technologies for identifying and utilizing mid-level radial convergence features in radar data suffer from problems such as discontinuous data boundaries, limited feature recognition capabilities, insufficient accuracy in convergence pair matching, and a lack of automated early warning mechanisms, resulting in insufficient accuracy and timeliness of gale warnings.
A circular filling strategy is adopted to stitch radar radial velocity data in a ring. Combined with an improved U-Net semantic segmentation network and channel attention mechanism, the convergence zone is semantically segmented. Convergence pair matching is performed through velocity-weighted centroid localization and spatial position constraints to achieve end-to-end convergence feature recognition and early warning level classification.
It improves the continuity and accuracy of convergence zone identification, enhances the ability to focus on key features, optimizes the matching accuracy of convergence pairs, and significantly improves the reliability and timeliness of gale warnings.
Smart Images
Figure CN122362390A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of meteorological radar signal processing and nowcasting technology, and more specifically, to a smart nowcasting system, method, and storage medium for strong winds based on radial velocity characteristics. Background Technology
[0002] Convective gales, due to their sudden onset and immense destructive power, often cause severe damage to buildings, transportation infrastructure, power lines, and agricultural production. Therefore, timely and accurate early warnings are of significant practical importance for disaster prevention and mitigation. In Doppler weather radar detection technology, mid-level radial convergence (MARC) is considered one of the key dynamic features indicating the impending occurrence of convective gales. The MARC feature is typically located in the mid-level of a convective storm and appears on the radar radial velocity map as a "positive-negative velocity pair" consisting of a continuously occurring "receding velocity zone" and an "approaching velocity zone" along the radial direction. Identifying this feature can provide an early warning time of approximately 10-30 minutes for gales such as downbursts.
[0003] Currently, the technical approaches for automatically extracting MARC features from radial velocity fields mainly fall into two categories. One category is based on traditional numerical calculation methods, such as using the two-dimensional Local Linear Least Squares Derivative (LLSD) method to calculate the radial velocity divergence shear, and then combining it with the Storm Cell Identification (SCIT) algorithm for the identification and voiding of three-dimensional convergence zones. The other category is based on image recognition methods, which convert radial velocity data into digital images, extract the contours of "positive-negative velocity" color blocks through multi-threshold binarization, and then locate convergence features through spatial correlation analysis.
[0004] However, the aforementioned existing technologies still have the following limitations when applied to actual business operations: Insufficient handling of spatial continuity in data: Radar data is organized in polar coordinates, resulting in natural numerical breaks at the 0° azimuth boundary. While existing technologies attempt to supplement the head and tail data using wrapping methods, simple data copying fails to fundamentally address the problem of deep learning models perceiving the loop-like topology during convolution, leading to the easy fragmentation or neglect of convergence features crossing the 0° boundary.
[0005] Limited feature recognition capabilities: While LLSD-based numerical algorithms can identify convergence zones that do not exhibit symmetrical "positive-negative velocity pairs," their recognition ability is highly dependent on preset physical thresholds, making them prone to missed detections in complex wind fields or weak convergence backgrounds. Image binarization-based methods, on the other hand, are limited in recognition accuracy by the degree of Gaussian blur and the selection of fixed thresholds, making it difficult to adaptively capture convergence zones with variable morphology and blurred boundaries. Furthermore, they are sensitive to isolated noise and ground clutter interference.
[0006] The accuracy of convergence pair matching needs improvement: Existing technologies typically use the geometric center of the profile or the outer rectangular frame to determine the spatial relationship between the recession zone and the approach zone. This approach does not fully consider the non-uniformity of wind speed distribution within the convergence zone, often leading to feature localization deviating from the core area with the highest actual convergence intensity, thus affecting the accuracy of convergence intensity calculation and vertical matching.
[0007] Insufficient depth of feature utilization: Most existing methods stop at detecting single elevation angles or three-dimensional MARC features, failing to effectively link the identified convergence features with subsequent early warning decisions. The generation of early warning indicators often relies on human experience, lacking an automated comprehensive evaluation and graded early warning mechanism for key parameters such as convergence intensity and vertical continuity.
[0008] In summary, overcoming the discontinuity of polar coordinate data boundaries, improving the robustness and accuracy of convergence zone identification in complex environments, achieving precise matching of convergence pairs, and ultimately forming an automated early warning classification system are technical problems that urgently need to be solved by those skilled in the art. Summary of the Invention
[0009] In view of this, the present invention proposes an intelligent early warning system, method and storage medium for near-term strong winds based on radial velocity characteristics, in order to solve the problems existing in the prior art.
[0010] To achieve the above objectives, this invention proposes a near-term strong wind intelligent early warning system based on radial velocity characteristics, comprising: The radar data preprocessing module is used to read weather radar base data, extract radial velocity field data at a specified elevation angle, and perform ring data stitching in the azimuth dimension to construct a radial velocity matrix. The semantic segmentation module is used to perform end-to-end convergence zone semantic segmentation on the radial velocity matrix using an improved U-Net semantic segmentation network to identify the recession velocity zone and the approach velocity zone. The instance segmentation module is used to perform connected component analysis on the semantic segmentation results, extract independent regression velocity region instances and approach velocity region instances, and calculate the geometric parameters and velocity statistics of each instance. The convergence pair matching module is used to match instances of the retreating velocity region with instances of the approaching velocity region based on spatial location and velocity gradient constraints, identifying mid-level radial convergence features; and The early warning decision module is used to calculate early warning indicators based on convergence intensity and vertical continuity, and to classify gale warning levels.
[0011] Furthermore, the radar data preprocessing module employs a cyclic filling strategy for ring-shaped stitching in the azimuth dimension, specifically including: Let the original radial velocity matrix be... Where A is the number of azimuth angles and R is the number of radial coulombs; Extend the matrix by k rows at the top and bottom, where Δθ is the azimuth resolution; Copy the last k rows of the original matrix and place them at the top of the extended matrix. Place the entire original matrix in the middle. Copy the first k rows of the original matrix and place them at the bottom of the extended matrix, forming a circular, continuous extended matrix. .
[0012] Furthermore, the improved U-Net network in the semantic segmentation module includes: The encoder consists of four downsampling stages, each of which includes two 3×3 convolutions, batch normalization, ReLU activation, and 2×2 max pooling. The decoder section adopts a symmetrical structure, upsampling through 2×2 transposed convolution, and splicing it with the corresponding layer of the encoder through skip connections; The channel attention module, embedded at the end of each stage of the encoder, uses the SENet structure to dynamically calibrate the feature maps; and the output layer maps the number of channels to three categories through 1×1 convolution, including background, recession velocity region, and proximity velocity region, and applies the Softmax function to obtain pixel-level classification probabilities.
[0013] Furthermore, the processing flow of the channel attention module includes: The channel descriptor is obtained by aggregating spatial information through global average pooling. :
[0014] .
[0015] The nonlinear relationship between channels is learned through two fully connected layers. The first fully connected layer reduces the number of channels to C / r and activates it with ReLU. The second fully connected layer restores the original number of channels C and generates normalized weights through the Sigmoid function. ; Multiply the weights s channel by channel by the original feature map to obtain the calibrated output. .
[0016] Furthermore, the semantic segmentation network uses a weighted sum of Dice loss and focus loss as its training objective function:
[0017] in, This represents the total loss during network training. for Loss measures the degree of overlap between the predicted segmentation and the ground truth annotation; It is a focus loss used to address class imbalance problems, focusing on samples that are difficult to classify.
[0018] Furthermore, the instance segmentation module includes: Each independent regression velocity region instance and approach velocity region instance is extracted using a connected component analysis algorithm; For instances in the receding velocity region, the 5th percentile of the velocity values of all pixels within it is calculated as the representative velocity; for instances in the approach velocity region, the 95th percentile is calculated as the representative velocity. If the absolute value of the velocity is lower than the preset threshold, it is determined to be a weak convergence region and is removed. The overlap relationship between adjacent instances is determined based on the intersection-union ratio of the bounding rectangle. Instances with an overlap of more than 70% are merged, and instances with larger absolute values of representative speeds are retained.
[0019] Furthermore, the matching algorithm of the convergence pair matching module includes: Traverse all instances of receding velocity zones, and use the center of each instance as a reference to search for spatially adjacent close velocity zone instances in the radial outward direction. The search range is limited to a fan-shaped area with a radial distance of less than 20 kilometers and an azimuth deviation of no more than 15 degrees. Calculate the difference between the representative velocity of the candidate approach velocity zone instance and the representative velocity of the current retreat velocity zone instance. If the absolute value of the velocity difference is greater than a preset threshold, a candidate convergence pair is initially formed. Calculate the Euclidean distance between the centroid of the instance in the receding velocity region and the centroid of the instance in the approach velocity region. If the distance is less than a preset maximum distance threshold, then retain the convergence pair. A non-maximum suppression strategy is used for deduplication. If the crossover ratio of the regression velocity region instances in two convergence pairs is greater than 80% and the crossover ratio of the close velocity region instances is also greater than 80%, then the one with the larger velocity difference is retained and the one with the smaller velocity difference is removed.
[0020] Furthermore, the centroid calculation employs a velocity-weighted average method, specifically including: For the backward speed range, the 90th percentile is taken as the baseline speed. For the near-speed range, the 10th percentile is used as the reference speed. ; Define the weight of each pixel as ; Calculate the weighted geometric center:
[0021] in, X , Y The coordinates of the centroid are the longitude and latitude coordinates. N This represents the total number of pixels within the wind speed zone instance. Let i be the weight of the i-th pixel. , Let be the longitude and latitude coordinates of the i-th pixel.
[0022] On the other hand, to achieve the above objectives, this invention proposes a near-term strong wind intelligent early warning method based on radial velocity characteristics, comprising the following steps: Step 1: Read the weather radar base data, extract the radial velocity field data at a specified elevation angle, and use a cyclic filling strategy to stitch the data in a ring along the azimuth dimension to construct a radial velocity matrix; Step 2: Input the radial velocity matrix into the improved U-Net semantic segmentation network to perform end-to-end convergence zone semantic segmentation and identify the recession velocity zone and the approach velocity zone; Step 3: Perform connected component analysis on the semantic segmentation results, extract independent regression velocity region instances and approach velocity region instances, and calculate the geometric parameters and velocity statistical features of each instance; Step 4: Based on spatial location and velocity gradient constraints, match instances of the recession velocity region with instances of the approach velocity region to identify mid-level radial convergence characteristics; Step 5: Calculate the early warning indicators based on the convergence intensity and vertical continuity, and classify the gale warning levels.
[0023] The present invention further proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described method for intelligent early warning of near-term strong winds based on radial velocity characteristics.
[0024] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention addresses the spatial discontinuity of polar coordinate data at the 0° boundary by employing a cyclic filling strategy to perform ring-shaped stitching of radar radial velocity data along the azimuth dimension. Combined with an improved U-Net network incorporating an embedded channel attention mechanism, it achieves end-to-end semantic segmentation of convergence zones. Furthermore, through velocity-weighted centroid localization and an intelligent matching algorithm for convergence pairs based on spatial location and velocity gradients, it accurately identifies mid-level radial convergence features. Finally, it classifies warning levels based on convergence intensity and vertical continuity. This invention effectively improves the continuity and accuracy of convergence zone identification, enhances the model's ability to focus on key features, and optimizes the matching accuracy of convergence pairs, thereby significantly improving the reliability and timeliness of gale warnings. Attached Figure Description
[0025] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1This is a schematic diagram of the near-term strong wind intelligent early warning system based on radial velocity characteristics according to the present invention. Figure 2 This is a schematic diagram of the process of the intelligent early warning method for near-term strong winds based on radial velocity characteristics according to the present invention. Detailed Implementation
[0026] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0027] This embodiment proposes a near-term strong wind intelligent early warning system based on radial velocity characteristics, such as... Figure 1 As shown, it includes: Step 1: Reading radar-based data and constructing the velocity field in polar coordinates The system first reads the base data file of a new-generation weather radar and extracts the radial velocity field data at a specified elevation angle. The radial velocity data is organized in polar coordinates, with azimuth as the first dimension and radial distance as the second dimension. Due to the radar scanning method causing data breaks at the 0° azimuth angle, this invention performs circular data stitching along the azimuth dimension to maintain the spatial continuity of the wind field. Specifically, a cyclic filling strategy is adopted, copying the data of the first 10° azimuth angle to the end of the matrix and copying the data of the last 10° azimuth angle to the beginning of the matrix to ensure a smooth transition along the azimuth dimension. Based on this, a radial velocity matrix is constructed as the input tensor for subsequent deep learning models.
[0028] This invention employs a cyclic filling strategy to perform ring-shaped stitching of radar radial velocity data along the azimuth dimension. Its core principle lies in resolving the spatial discontinuity problem of polar coordinate data at the 0° azimuth angle. Because radar scanning methods result in 0° and 360° being physically adjacent, but located in the first and last rows of a matrix representation, respectively, if directly used as input to a deep learning model, the model struggles to capture the wind field structure crossing this boundary, thus affecting the accuracy of convergence zone identification.
[0029] The specific implementation steps are as follows: Let the original radial velocity matrix be...
[0030] Where A represents the number of azimuth angles and R represents the number of radial columns. To maintain the circumferential continuity of the azimuth dimension, k rows are added at the top and bottom of the matrix, where k corresponds to the initial angle range; in this embodiment, k is taken as 10°. The number of added rows is calculated based on the azimuth resolution Δθ. .
[0031] The loop filling operation is as follows: Copy the last k rows of the original matrix (corresponding to azimuth angles 360° - kΔθ to 360°) and place them on top of the extended matrix; Place the original matrix entirely in the middle of the extended matrix; Copy the first k rows of the original matrix (corresponding to azimuth angles from 0° to kΔθ) and place them at the bottom of the extended matrix.
[0032] The specific formula is as follows: Construct the extended matrix
[0033]
[0034] Where is the row index and is the column index. After this processing, the extended matrix forms a continuous circular data stream along the azimuth dimension, explicitly encoding the adjacency relationship between 0° and 360°.
[0035] Finally As the input tensor of the deep learning model, the shape is (A+2k, R). This strategy ensures that the model can make full use of the ring-shaped topology in the azimuth dimension, improve the perception of convergence features that cross the 0° boundary, and avoid the feature extraction bias caused by data fragmentation in traditional methods.
[0036] Step 2: Multi-scale radial velocity field feature extraction based on attention mechanism Unlike traditional methods based on fixed threshold binarization and contour search, this invention introduces an improved U-Net semantic segmentation network, combined with a channel attention mechanism, to perform end-to-end semantic segmentation of the radial velocity matrix's convergence regions. Specifically, a deep neural network with an encoder-decoder structure is constructed. The encoder part uses a residual network to extract multi-scale features, and convolutional block attention modules are embedded at skip connections to enhance the network's responsiveness to coupled regions of positive and negative velocities. The network input is a preprocessed radial velocity matrix, and the output is a semantic segmentation map of the same size as the input, with pixel categories including background, recession velocity region, and proximity velocity region.
[0037] Network structure design: The input is a radial velocity matrix that has been cyclically filled. ,in To extend the rear azimuth dimension, WThe number of channels is the radial number of classes. The encoder consists of four downsampling stages, each containing two 3×3 convolutions followed by batch normalization and ReLU activation, then halving the feature map size through 2×2 max pooling. A channel attention module is embedded at the end of each stage to dynamically calibrate the extracted feature maps. The decoder employs a symmetrical structure, upsampling through 2×2 transposed convolutions and concatenating them with the corresponding encoder layers via skip connections. The concatenated feature map is then refined through two 3×3 convolutions. Finally, a 1×1 convolution maps the number of channels to the number of classes. The data includes the background, the recession velocity region, and the approach velocity region, and a Softmax function is applied to obtain pixel-level classification probabilities.
[0038] Channel attention mechanism: This invention selects the SENet module as the channel attention implementation method. For the feature map output by the encoder at a certain stage, spatial information is first aggregated through global average pooling to obtain the channel descriptor. :
[0039] Subsequently, the nonlinear relationship between channels is learned through two fully connected layers: the first fully connected layer reduces the number of channels to C / r (compression ratio). r =16), activated by ReLU; the second fully connected layer restores the original number of channels C, and generates normalized weights through the Sigmoid function. :
[0040] in , The weights are learnable weights, δ is ReLU, and σ is Sigmoid. Multiplying the weights s channel-by-channel by the original feature map yields the calibrated output:
[0041] This mechanism enables the network to automatically enhance its focus on key features of the convergence zone and suppress irrelevant noise.
[0042] During the model's training phase, radar data of historical strong wind weather events and their manually labeled convergence zones are used as supervision signals. To address the class imbalance problem, a weighted sum of Dice loss and focus loss is employed:
[0043] in, This represents the total loss during network training. for Loss measures the degree of overlap between the predicted segmentation and the ground truth annotation; This is a focus loss function used to address class imbalance by focusing on hard-to-classify samples. As a balance coefficient, in this embodiment, we take... .
[0044] Through end-to-end training, the network can accurately segment the convergence region in the radial velocity map, providing a reliable basis for subsequent gale warnings.
[0045] Step 3: Convergence Zone Instance Segmentation and Geometric Parameter Calculation After obtaining the semantic segmentation results, the system uses a connected component analysis algorithm to extract each independent regression velocity region instance and proximity velocity region instance. For each instance, the geometric center, bounding rectangle, area, and velocity statistical feature parameters are calculated. To remove isolated noise regions, a statistical filtering strategy based on velocity value distribution is introduced: for each regression velocity region instance, the 5th percentile of the velocity values of all pixels within it is calculated as the representative velocity; for the proximity velocity region, its 95th percentile is calculated as the representative velocity. If the absolute value of the representative velocity is lower than a preset threshold, it is determined to be a weak convergence region and is removed. Simultaneously, the overlap relationship between adjacent instances is determined based on the intersection-union ratio of the bounding rectangle; instances with an overlap exceeding 70% are merged, and instances with larger absolute values of representative velocities are retained.
[0046] Step 4: Intelligent Matching and Spatial Correlation Analysis of Convergence Pairs To accurately identify mid-level radial convergence features, this invention proposes a convergence pair matching algorithm based on spatial position and velocity gradient constraints. The specific steps are as follows: First, all regression velocity zone instances are traversed. Using the center of each instance as a reference, adjacent approach velocity zone instances are searched radially outward. The search area is limited to a fan-shaped region with a radial distance of no more than 20 kilometers and an azimuth deviation of no more than 15 degrees. The difference between the representative velocity of the candidate approach velocity zone instance and the representative velocity of the current regression velocity zone instance is calculated. If the absolute value of the velocity difference is greater than a preset threshold, a preliminary candidate convergence pair is formed.
[0047] Secondly, the Euclidean distance between the centroid of the receding velocity region instance and the centroid of the approach velocity region instance is calculated. If the distance is less than the preset maximum distance threshold, the convergence pair is retained.
[0048] Traditional geometric centering only considers pixel coordinates, ignoring the uneven distribution of wind speed values, which tends to bias the centroid towards the center of the region rather than the core area with the strongest convergence. The velocity-weighted averaging method assigns higher weights to pixels with extreme velocity values, shifting the centroid towards the area with the strongest convergence, thereby improving the accuracy of subsequent convergence pair matching and distance calculation. Therefore, this embodiment uses the velocity-weighted averaging method for centroid calculation. That is, for all pixels within the instance, the absolute value of the difference between their velocity value and the reference velocity is used as the weight to calculate the weighted geometric center, improving the accuracy of centroid positioning. The specific calculation steps include: Determine the baseline velocity: Suppose a wind speed zone instance contains N pixels, and the velocity value of each pixel is [value]. For the recession velocity zone, take the 90th quantile as the baseline; for the approach velocity zone, take the 10th quantile as the baseline. The quantile threshold can effectively filter outliers, ensuring that the baseline represents extreme value levels.
[0049] To avoid zero-weight pixels being completely ignored and to maintain numerical stability, for each pixel... i Define its weights Add 1 to the absolute value of the difference between the speed value and the reference value, that is:
[0050] Let i be the weight of the i-th pixel. Let be the radial velocity value of the i-th pixel. The base speed for this wind speed zone is taken as the 90th percentile for the retreat zone and the 10th percentile for the approach zone.
[0051] Let the latitude and longitude coordinates of pixel i be ( , Then, the formula for calculating the centroid coordinates (X,Y) is:
[0052] in, X , Y The coordinates of the centroid are the longitude and latitude coordinates. N This represents the total number of pixels within the wind speed zone instance. Let i be the weight of the i-th pixel. , Let be the longitude and latitude coordinates of the i-th pixel.
[0053] Finally, for multiple overlapping convergence pairs that may exist in the same region, a non-maximum suppression strategy is used to remove duplicates. The specific judgment criterion is as follows: if the cross-union ratio of the regression velocity region instances in two convergence pairs is greater than 80%, and the cross-union ratio of the adjacent velocity region instances is also greater than 80%, then they are considered redundant convergence pairs. The pair with the larger velocity difference is retained, and the pair with the smaller velocity difference is discarded.
[0054] Step 5: Calculation and Classification of Gale Warning Indicators For each valid convergence pair ultimately retained, the system treats each valid convergence pair as a potential mid-level radial convergence feature and further calculates early warning indicators. These indicators include convergence intensity, vertical thickness of the convergence zone, and the altitude of the convergence zone. Convergence intensity is defined as the absolute value of the difference between the representative velocity of the recession velocity zone and the representative velocity of the approach velocity zone. The vertical thickness of the convergence zone is obtained by matching convergence pairs at adjacent elevation angles. If convergence pairs are detected at the same location at two or more consecutive elevation angles, vertical continuity is determined, increasing the confidence level of the early warning.
[0055] Based on convergence intensity and vertical continuity, gale warning levels are divided into three levels: a red warning is issued if the convergence intensity is greater than or equal to 10 m / s and has vertical continuity; an orange warning is issued if the convergence intensity is between 8 m / s and 10 m / s, or greater than 10 m / s but lacks vertical continuity; and a yellow warning is issued if the convergence intensity is between 6 m / s and 8 m / s. Warning information is displayed graphically in the system interface and supports automatic push to relevant business platforms.
[0056] Furthermore, to adapt to the differences in convective storm characteristics across different regions and seasons, this embodiment designs an online learning and dynamic update mechanism for the model. In daily scheduled tasks, the system compares the automatically identified results from the previous day with manually reviewed real-time wind records, extracting false positives and false negatives to construct an incremental training dataset. Incremental learning algorithms are then used to fine-tune the semantic segmentation network, achieving continuous optimization and adaptation of the model.
[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A near-term strong wind intelligent early warning system based on radial velocity characteristics, characterized in that, include: The radar data preprocessing module is used to read weather radar base data, extract radial velocity field data at a specified elevation angle, and perform ring data stitching in the azimuth dimension to construct a radial velocity matrix. The semantic segmentation module is used to perform end-to-end convergence zone semantic segmentation on the radial velocity matrix using an improved U-Net semantic segmentation network to identify the recession velocity zone and the approach velocity zone. The instance segmentation module is used to perform connected component analysis on the semantic segmentation results, extract independent regression velocity region instances and approach velocity region instances, and calculate the geometric parameters and velocity statistics of each instance. The convergence pair matching module is used to match instances of the receding velocity region with instances of the approach velocity region based on spatial location and velocity gradient constraints, and to identify mid-level radial convergence features. as well as The early warning decision module is used to calculate early warning indicators based on convergence intensity and vertical continuity, and to classify gale warning levels.
2. The system according to claim 1, characterized in that, The radar data preprocessing module employs a cyclic filling strategy for ring-shaped stitching along the azimuth dimension, specifically including: Let the original radial velocity matrix be... Where A is the number of azimuth angles and R is the number of radial coulombs; Extend the matrix by k rows at the top and bottom, where Δθ is the azimuth resolution; Copy the last k rows of the original matrix and place them at the top of the extended matrix. Place the entire original matrix in the middle. Copy the first k rows of the original matrix and place them at the bottom of the extended matrix, forming a circular, continuous extended matrix. .
3. The system according to claim 1, characterized in that, The improved U-Net network in the semantic segmentation module includes: The encoder consists of four downsampling stages, each of which includes two 3×3 convolutions, batch normalization, ReLU activation, and 2×2 max pooling. The decoder section adopts a symmetrical structure, upsampling through 2×2 transposed convolution, and splicing it with the corresponding layer of the encoder through skip connections; The channel attention module, embedded at the end of each stage of the encoder, uses the SENet structure to dynamically calibrate the feature maps; and the output layer maps the number of channels to three categories through 1×1 convolution, including background, recession velocity region, and proximity velocity region, and applies the Softmax function to obtain pixel-level classification probabilities.
4. The system according to claim 3, characterized in that, The processing flow of the channel attention module includes: The channel descriptor is obtained by aggregating spatial information through global average pooling. : ； ； The nonlinear relationship between channels is learned through two fully connected layers. The first fully connected layer reduces the number of channels to C / r and activates it with ReLU. The second fully connected layer restores the original number of channels C and generates normalized weights through the Sigmoid function. ; Multiply the weights s channel by channel by the original feature map to obtain the calibrated output. .
5. The system according to claim 3, characterized in that, The semantic segmentation network uses a weighted sum of Dice loss and focus loss as its training objective function: ； in, This represents the total loss during network training. for Loss measures the degree of overlap between the predicted segmentation and the ground truth annotation; It is a focus loss used to address class imbalance problems, focusing on samples that are difficult to classify.
6. The system according to claim 1, characterized in that, The instance segmentation module includes: Each independent regression velocity region instance and approach velocity region instance is extracted using a connected component analysis algorithm; For instances in the receding velocity region, the 5th percentile of the velocity values of all pixels within it is calculated as the representative velocity; for instances in the approach velocity region, the 95th percentile is calculated as the representative velocity. If the absolute value of the velocity is lower than the preset threshold, it is determined to be a weak convergence region and is removed. The overlap relationship between adjacent instances is determined based on the intersection-union ratio of the bounding rectangle. Instances with an overlap of more than 70% are merged, and instances with larger absolute values of representative speeds are retained.
7. The system according to claim 1, characterized in that, The matching algorithm of the convergence pair matching module includes: Traverse all instances of receding velocity zones, and use the center of each instance as a reference to search for spatially adjacent close velocity zone instances in the radial outward direction. The search range is limited to a fan-shaped area with a radial distance of less than 20 kilometers and an azimuth deviation of no more than 15 degrees. Calculate the difference between the representative velocity of the candidate approach velocity zone instance and the representative velocity of the current retreat velocity zone instance. If the absolute value of the velocity difference is greater than a preset threshold, a candidate convergence pair is initially formed. Calculate the Euclidean distance between the centroid of the instance in the recession velocity region and the centroid of the instance in the approach velocity region. If this distance is less than a preset maximum distance threshold, retain the convergence pair; and A non-maximum suppression strategy is used for deduplication. If the crossover ratio of the regression velocity region instances in two convergence pairs is greater than 80% and the crossover ratio of the close velocity region instances is also greater than 80%, then the one with the larger velocity difference is retained and the one with the smaller velocity difference is removed.
8. The system according to claim 7, characterized in that, The centroid calculation employs a velocity-weighted average method, specifically including: For the backward speed range, the 90th percentile is taken as the baseline speed. For the near-speed range, the 10th percentile is used as the reference speed. ; Define the weight of each pixel as ; Calculate the weighted geometric center: ; in, X , Y The coordinates of the centroid are the longitude and latitude coordinates. N This represents the total number of pixels within the wind speed zone instance. Let i be the weight of the i-th pixel. , Let be the longitude and latitude coordinates of the i-th pixel.
9. A method for intelligent early warning of near-term strong winds based on radial velocity characteristics, characterized in that, Includes the following steps: Step 1: Read the weather radar base data, extract the radial velocity field data at a specified elevation angle, and use a cyclic filling strategy to stitch the data in a ring along the azimuth dimension to construct a radial velocity matrix; Step 2: Input the radial velocity matrix into the improved U-Net semantic segmentation network to perform end-to-end convergence zone semantic segmentation and identify the recession velocity zone and the approach velocity zone; Step 3: Perform connected component analysis on the semantic segmentation results, extract independent regression velocity region instances and approach velocity region instances, and calculate the geometric parameters and velocity statistical features of each instance; Step 4: Based on spatial location and velocity gradient constraints, match instances of the recession velocity region with instances of the approach velocity region to identify mid-level radial convergence characteristics; Step 5: Calculate the early warning indicators based on the convergence intensity and vertical continuity, and classify the gale warning levels.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the intelligent early warning method for near-term strong winds based on radial velocity characteristics as described in claim 9.