Hook-shaped echo identification method, apparatus, equipment and storage medium

Abstract

This application discloses a method, apparatus, device, and storage medium for identifying hook echoes. The method includes: acquiring historical weather radar echo images; identifying hook echoes in the historical weather radar echo images to obtain radar echo training images; training a deep convolutional neural network model based on the radar echo training images to obtain a trained deep convolutional neural network model; acquiring an echo image to be identified; and using the trained deep convolutional neural network model to identify hook echoes in the echo image to be identified. This solution introduces an automatic feature extraction and pattern recognition mechanism based on deep learning, enabling the model to autonomously learn the feature representations of hook echoes in terms of morphological structure, reflectivity distribution, and spatial continuity from large-scale radar echo data. This avoids the problems of strong reliance on experience and insufficient generalization ability of traditional manual threshold setting and rule matching methods.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a hook-shaped echo recognition method, apparatus, device, and storage medium. Background Technology

[0002] With the continuous development of severe convective weather monitoring and disaster early warning services, hook echoes, as a typical "quasi-6-shaped" structural feature of supercells in the reflectivity factor field of weather radar, have become a key criterion with important indicative significance in gale and tornado warnings. Hook echoes usually reflect the presence of strong rotating convective structures such as mesocyclones, and their appearance is often accompanied by significant risks of severe weather. Therefore, timely and accurate identification of hook echoes is of fundamental importance for improving the monitoring and early warning capabilities for severe weather.

[0003] However, in practical meteorological operations, the identification of hook echoes has long relied on manual visual interpretation. This not only places a high degree of dependence on the experience of forecasters but also makes it difficult to meet the needs of real-time monitoring and large-scale automated processing, especially given the complex and variable echo morphology and high timeliness requirements of operations. Existing intelligent hook echo identification methods are mostly based on predefined standard templates, determining the presence of hook echoes by calculating the similarity between radar echoes and templates. However, due to significant differences in scale, curvature, and reflectivity distribution of hook echoes under different weather backgrounds and at different development stages, the aforementioned template matching methods generally suffer from poor adaptability and low identification accuracy, making it difficult to meet the practical application requirements of meteorological operations. Summary of the Invention

[0004] This application provides a method, apparatus, device, and storage medium for hook echo identification, which enables automated and accurate identification of hook echoes in meteorological radar echoes. This application avoids the problems of rigid rules and reliance on experience in traditional template matching or manual interpretation methods, significantly improving the accuracy, stability, and processing efficiency of hook echo identification. It is suitable for meteorological operational scenarios with high requirements for identification accuracy and timeliness, such as severe convective weather monitoring and disaster early warning, and provides effective support for building an intelligent and scalable hook echo identification system.

[0005] In a first aspect, this application provides a hook-shaped echo identification method, comprising:

[0006] Acquire historical weather radar echo images, identify hook-shaped echoes in the historical weather radar echo images, and obtain radar echo training images.

[0007] The deep convolutional neural network model is trained based on the radar echo training image to obtain the trained deep convolutional neural network model.

[0008] Obtain the echo image to be identified, and use the trained deep convolutional neural network model to identify the hook-shaped echo in the echo image.

[0009] Secondly, this application provides a hook-shaped echo identification device, comprising:

[0010] The data identification module is used to acquire historical weather radar echo images, identify the hook-shaped echoes in the historical weather radar echo images, and obtain radar echo training images.

[0011] The model training module is used to train the deep convolutional neural network model based on the radar echo training image to obtain the trained deep convolutional neural network model.

[0012] The echo recognition module is used to acquire the echo image to be recognized and to perform hook echo recognition on the echo image to be recognized using the trained deep convolutional neural network model.

[0013] Thirdly, this application provides a hook-shaped echo identification device, comprising:

[0014] One or more processors;

[0015] A memory that stores one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the hook-echo recognition method as described in the first aspect.

[0016] Fourthly, this application provides a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform the hook-echo recognition method as described in the first aspect.

[0017] In this application, a systematic acquisition and preprocessing of historical meteorological radar echo images is performed, and echo regions exhibiting hook-shaped structural features are finely labeled to construct a radar echo training image set containing prior information about hook echoes. This forms the learning foundation for hook echo recognition oriented towards meteorological operational semantics. During the model building phase, a deep convolutional neural network model is trained based on the radar echo training images. This allows the model to automatically learn the intrinsic relationships between hook echoes in multiple dimensions, such as spatial morphology, structural continuity, and reflectivity distribution, completing the mapping from raw echo data to high-level semantic feature representation. In the hook echo recognition execution phase, the echo image to be recognized is acquired and input into the trained deep convolutional neural network model. The model performs multi-level feature extraction and discrimination calculations on the features in the echo image, generating a recognition result reflecting the degree of matching between the echo to be recognized and the hook echo features. When the discrimination result meets the preset recognition conditions, the recognition and localization of the hook echo are automatically completed. This method effectively overcomes the problems of traditional template matching methods, such as strong dependence on fixed shapes and insufficient adaptability, through training sample-driven feature learning and model inference mechanisms. It significantly improves the accuracy, stability and real-time processing capability of hook echo identification under complex weather backgrounds and diverse echo shapes, providing reliable and scalable technical support for severe convective weather monitoring and disaster early warning. Attached Figure Description

[0018] Figure 1 This is a flowchart of a hook-shaped echo recognition method provided in an embodiment of this application;

[0019] Figure 2 This is a flowchart of the training data augmentation method provided in the embodiments of this application;

[0020] Figure 3 This is a flowchart of a model training method based on the SiLU activation function provided in an embodiment of this application;

[0021] Figure 4 This is a flowchart of a model training method based on the BCE loss function provided in an embodiment of this application;

[0022] Figure 5 This is a flowchart of a model training method based on the DFL loss function provided in an embodiment of this application;

[0023] Figure 6 This is a flowchart of the model training method based on the CIoU loss function provided in the embodiments of this application;

[0024] Figure 7 This is a flowchart of the model verification method provided in the embodiments of this application;

[0025] Figure 8This is a schematic diagram of the hook-shaped echo recognition device provided in the embodiments of this application;

[0026] Figure 9 This is a schematic diagram of the structure of a hook-shaped echo recognition device provided in an embodiment of this application. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, specific embodiments of this application will be described in further detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely for explaining this application and not for limiting it. It should also be noted that, for ease of description, only the parts relevant to this application are shown in the drawings, not all of them. Before discussing exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe operations (or steps) as being processed sequentially, many of these operations can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. A process can be terminated when its operation is completed, but it may also have additional steps not included in the drawings. A process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0028] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0029] In current meteorological operational systems, hook echoes are a key identification target in the monitoring and early warning of severe convective weather, playing a significant indicative role in gale and tornado warnings. Hook echoes typically exhibit a "6"-shaped echo structure, representing a typical image morphology of supercells in the reflectivity factor field of weather radar. Their appearance is often closely related to strong rotating convective systems such as mesocyclones. Therefore, accurate identification of hook echoes is a crucial technical foundation for improving the ability to issue severe weather warnings. However, with the expansion of meteorological radar observation range and the continuous increase in data update frequency, the scale and morphological complexity of radar echo data continue to grow. Existing hook echo identification methods are gradually revealing significant shortcomings in terms of real-time performance and automation, making it difficult to meet the actual needs of operational systems for rapid response and highly reliable early warnings.

[0030] Existing hook echo identification technologies mainly fall into two categories: those based on manual interpretation and those based on fixed template matching. Traditional manual identification methods rely on forecasters' visual analysis and experience in radar echo images, which is not only highly subjective and labor-intensive, but also struggles to guarantee consistency and timeliness in large-scale, multi-time continuous monitoring scenarios. Hook echo identification schemes based on standard templates typically involve pre-constructing hook echo templates and calculating the similarity between the target echo and the template to complete the identification. However, this type of scheme is poorly adaptable to variations in the scale, curvature, and reflectivity distribution of hook echoes. When hook echo morphologies differ significantly under different weather conditions and at different development stages, template matching results are prone to missed or false detections, and the overall identification accuracy and stability fail to meet the practical application requirements of meteorological operations. To address the aforementioned issues, this invention introduces deep learning technology. By training a large number of historical radar echo samples, the model can automatically learn a high-dimensional representation of the spatial structure and morphological features of hook echoes, enabling efficient, rapid, and accurate identification and localization of hook echoes. This can effectively support the monitoring and early warning applications of strong winds and tornadoes in meteorological operations.

[0031] To address the aforementioned issues, this embodiment provides a hook-shaped echo recognition method. Through a training data-driven feature learning and model inference mechanism, it achieves efficient and stable recognition of hook-shaped echoes. The method first acquires historical meteorological radar echo images and systematically identifies echo regions with hook-shaped structural features, constructing a radar echo training image set with clear semantic labels, thus establishing a reliable sample foundation for model learning. During the model training phase, a deep convolutional neural network model is trained based on the radar echo training images. This enables the model to automatically learn the discriminative features of hook-shaped echoes in terms of spatial morphology, structural continuity, and reflectivity distribution during multi-layer feature extraction and nonlinear mapping, forming a stable recognition capability. In the recognition execution phase, the echo image to be recognized is acquired and input into the trained deep convolutional neural network model. The model performs end-to-end feature analysis and discriminative calculations on the echo image, outputting the hook-shaped echo recognition result. This solution avoids the problems of rigid rules and strong template dependence in traditional methods by using model training and inference processes based on historical labeled samples. While ensuring that the recognition logic is clear and controllable, it significantly improves the accuracy, robustness and real-time processing capability of hook echo recognition. It is suitable for meteorological business scenarios with high requirements for recognition efficiency and reliability, such as severe convective weather monitoring and disaster early warning.

[0032] Figure 1 A flowchart of a hook-shaped echo recognition method provided in an embodiment of this application is given. (Reference) Figure 1 The hook-shaped echo identification method specifically includes:

[0033] S110. Acquire historical weather radar echo images, identify the hook-shaped echoes in the historical weather radar echo images, and obtain radar echo training images.

[0034] In some embodiments, historical weather radar echo images can be a collection of radar echo image data acquired by weather radar under different time periods and weather conditions. Hook echoes are characteristic echo structures in radar echo images that exhibit a hook-like or curved shape, used to characterize strong convective or rotating weather systems. Radar echo training images can be sample images containing hook echo identification information, used for subsequent model training or recognition rule construction. First, historical weather radar echo images are acquired from a historical radar data storage system or meteorological data platform; then, the historical weather radar echo images are analyzed and processed, and hook echo regions are identified in the images, thereby generating corresponding radar echo training images.

[0035] In one embodiment, the hook echo can be identified by having meteorological experts manually interpret historical weather radar echo images and mark the region outline, center position, or pixel range corresponding to the hook echo in the image, thus forming a training sample with manually labeled information.

[0036] In one embodiment, the hook echo can be identified by automatically or semi-automatically detecting historical weather radar echo images based on existing hook echo identification rules or initial identification models, and confirming and identifying the detected hook echo regions to improve the efficiency of training sample construction.

[0037] In one embodiment, the radar echo training image may include pixel annotation information, category labels, or confidence information of the hook-shaped echo region, which can be used to guide the model to learn the spatial morphological features and echo distribution features of the hook-shaped echo during subsequent training.

[0038] S120. The deep convolutional neural network model is trained based on the radar echo training image to obtain the trained deep convolutional neural network model.

[0039] In some embodiments, the deep convolutional neural network model can be a deep learning model for automatically extracting spatial morphological features and performing pattern recognition from radar echo images. The radar echo training images can be labeled sample data in historical meteorological radar echo images after hook echo identification. The trained deep convolutional neural network model can be a target model with hook echo recognition capabilities. Based on the obtained radar echo training images, the pre-constructed deep convolutional neural network model is trained to learn the spatial structure features, morphological distribution features, and intensity variation features of hook echoes in radar echo images, thereby obtaining the trained deep convolutional neural network model.

[0040] In one embodiment, the deep convolutional neural network model can be trained by using radar echo training images as model inputs, using the corresponding hook-shaped echo identification information as supervision labels, and continuously updating the model parameters through forward and backward propagation processes to minimize the error between the model output and the identification results.

[0041] In one embodiment, preprocessing operations, including image normalization, scale adjustment, or data augmentation, can be performed on the radar echo training images during training to improve the model's generalization ability to hook echo features of different scales and intensities.

[0042] In one embodiment, the training process of a deep convolutional neural network model can employ a multi-round iterative training strategy, and the model performance can be evaluated using a validation set during the training process. When the model's recognition accuracy or loss function meets preset conditions, the training ends and the trained deep convolutional neural network model is output.

[0043] Optionally, Figure 2 A flowchart of the training data augmentation method provided in this application embodiment is given. The radar echo training image includes a first training image with hook-shaped echoes and a second training image without hook-shaped echoes, as referenced. Figure 2 The training data augmentation method specifically includes:

[0044] S1201. Rotate the first training image at preset angle intervals to obtain the third training image.

[0045] For example, the first training image can be a radar echo training image containing hook-shaped echo identification information, the preset angle interval can be an angle parameter used to control the rotation amplitude of the image, and the third training image can be an enhanced training sample generated by rotation transformation based on the first training image. The first training image is rotated according to the preset angle interval, so that the first training image is rotated around the image center or a preset reference point by multiple different angles, thereby generating the corresponding third training image, which is used to expand the number of training samples and enhance sample diversity.

[0046] In one embodiment, the preset angle interval can be a fixed angle step, such as 5°, 10° or other preset angle values. By rotating the first training image at multiple angles, the presentation of the hook-shaped echo in different orientations or different echo patterns can be simulated.

[0047] In one embodiment, during the rotation of the first training image, the identification information of the hook echo is simultaneously rotated and mapped by the same angle to ensure the consistency of the third training image and its corresponding identification information in spatial position.

[0048] In one embodiment, the third training image generated by the above rotation processing together with the original first training image constitutes an expanded training sample set, which helps to improve the robustness and generalization ability of the deep convolutional neural network model to changes in the direction of the hook echo, morphological rotation and spatial distribution differences.

[0049] S1202. The deep convolutional neural network model is trained using the first training image, the second training image, and the third training image to obtain a trained deep convolutional neural network model.

[0050] For example, the first training image can be the original training sample obtained after identifying hook echoes in historical weather radar echo images; the second training image can be an extended training sample generated based on the first training image using a preset data augmentation method; and the third training image can be a training sample obtained after rotating the first training image at preset angular intervals. The deep convolutional neural network model can be used to automatically extract features from radar echo images and complete hook echo identification. The first, second, and third training images are input together as a training dataset into the deep convolutional neural network model for joint training. This allows the model to learn the spatial morphological features, directional change features, and echo distribution features of hook echoes under diverse sample conditions, thereby obtaining the trained deep convolutional neural network model.

[0051] In one embodiment, the deep convolutional neural network model can be trained by using the hook-shaped echo label information corresponding to the first training image, the second training image, and the third training image as supervision labels, and iteratively updating the model parameters through forward and backward propagation processes to minimize the error between the model output and the labeling results.

[0052] In one embodiment, during the training process, the training data can be divided into a training subset and a validation subset according to a preset ratio, and the recognition accuracy, loss function change trend or convergence of the model can be evaluated based on the validation subset. When the evaluation result meets the preset conditions, the training ends and the trained deep convolutional neural network model is output.

[0053] In one embodiment, by introducing a joint training strategy of the first training image, the second training image, and the third training image, the ability of the deep convolutional neural network model to recognize hook echoes under different scales, directions, and morphological conditions can be effectively improved, thereby enhancing the generalization performance and stability of the model in actual meteorological radar echo recognition scenarios.

[0054] Optionally, Figure 3A flowchart of the model training method based on the SiLU activation function provided in this application is given. (Reference) Figure 3 The model training method based on the SiLU activation function specifically includes:

[0055] S1203. Input the radar echo training image into the input layer of the deep convolutional neural network model.

[0056] For example, the radar echo training image can be a training sample image containing hook-shaped echo identification information, and the input layer of the deep convolutional neural network model can be the initial network layer of the model used to receive and carry image data, introducing external image data into the model computation process. By inputting the radar echo training image into the input layer of the deep convolutional neural network model, the training image serves as the model's raw input data for subsequent feature extraction and parameter learning processes.

[0057] In one embodiment, before inputting radar echo training images into the input layer, preprocessing operations may be performed on the training images, including image size normalization, pixel value normalization, or channel adjustment, to match the data format of the training images to the input requirements of the deep convolutional neural network model.

[0058] In one embodiment, radar echo training images can be input into the input layer of a deep convolutional neural network model in batches to improve computational efficiency during training and support batch updates of model parameters.

[0059] Through the above steps, radar echo training images can be effectively introduced into the deep convolutional neural network model, laying a data foundation for the subsequent learning and extraction of hook echo features by convolutional layers, pooling layers, and fully connected layers.

[0060] S1204. The radar echo training image is subjected to feature extraction through the convolutional layer of the deep convolutional neural network model to obtain the first feature map.

[0061] For example, the convolutional layer of a deep convolutional neural network model can be a network layer structure used to perform local receptive field calculations and weight sharing operations on the input image. The radar echo training image can be the training sample data input to the model's input layer. The first feature map can be an intermediate feature representation generated by the convolutional layer performing convolution operations on the radar echo training image, used to characterize the local spatial features in the radar echo image. The radar echo training image input to the deep convolutional neural network model is passed to the convolutional layer. Through the sliding calculation of the convolutional kernel in the image space, the local texture features, edge features, and morphological features in the radar echo training image are extracted, thereby obtaining the corresponding first feature map.

[0062] In one embodiment, the convolutional layer may include multiple convolutional kernels of different sizes or with different weight parameters. Each convolutional kernel is used to extract a specific feature pattern in the radar echo training image. Multiple convolutional kernels generate multiple first feature maps to characterize the structural information in the radar echo image from different feature dimensions.

[0063] In one embodiment, a nonlinear activation function can be combined with the convolution operation to process the convolution result, so that the first feature map can more effectively express the nonlinear feature distribution of the hook echo in the radar echo training image.

[0064] Through the above steps, the original radar echo training image can be converted into a first feature map containing rich spatial feature information, which provides a foundation for the subsequent network layers to further extract and recognize the high-level semantic features of the hook echo.

[0065] S1205. Perform pooling processing on the first feature map to obtain the second feature map.

[0066] For example, pooling can be a downsampling operation performed on a feature map to reduce its spatial resolution and computational complexity while preserving key feature information. The first feature map can be an intermediate feature representation output by a convolutional layer, and the second feature map can be the feature map result obtained after pooling. Pooling is performed on the first feature map, and the corresponding second feature map is obtained by aggregating feature values ​​within a preset pooling window, thereby achieving scale compression and enhanced feature robustness.

[0067] In one embodiment, pooling processing may include max pooling, average pooling, or weighted pooling, and by selecting the maximum value, average value, or weighted result within the pooling window, representative feature responses in the first feature map are retained.

[0068] In one embodiment, pooling can reduce the sensitivity of the first feature map to local noise and small displacement changes, allowing the second feature map to focus more on the overall morphological features of the hook-shaped echo in the radar echo image.

[0069] By taking the above steps, we can reduce the dimension of the feature map and the amount of computation while retaining the key information that is discriminative for hook echo recognition, providing stable feature input for feature learning and classification decisions in subsequent network layers.

[0070] S1206. Activate the second feature map using the SiLU activation function to obtain the third feature map.

[0071] For example, the SiLU activation function can be a continuously differentiable nonlinear activation function used to perform nonlinear mapping on the values ​​in the feature map. The second feature map can be the feature representation obtained after pooling, and the third feature map can be the output feature map obtained after performing activation processing on the second feature map. By using the SiLU activation function to perform activation processing on each feature value in the second feature map, and performing nonlinear transformation on the feature values, the corresponding third feature map is obtained, giving the feature representation a stronger nonlinear expressive power.

[0072] In one embodiment, the SiLU activation function can be expressed as:

[0073]

[0074] in, This represents the eigenvalues ​​in the second feature map. This represents the output of the Sigmoid function. This function's form ensures that eigenvalues ​​grow smoothly in the positive interval and maintain a certain gradient in the negative interval, thus avoiding the vanishing gradient problem.

[0075] In one embodiment, by introducing the SiLU activation function, the deep convolutional neural network model can be enhanced to model the complex morphological features of hook-shaped echoes in radar echo images, thereby improving the convergence stability and recognition accuracy of the model during training and inference.

[0076] Through the above steps, the second feature map can be mapped into a third feature map with richer nonlinear feature expression, providing effective input for subsequent network layers to further extract high-level semantic features.

[0077] S1207. Based on the third feature map, generate the category prediction result and spatial positioning result of the hook echo through the output layer.

[0078] For example, the output layer can be a network layer structure set at the end of a deep convolutional neural network model, used to comprehensively analyze the third feature map and generate the final recognition result; the third feature map can be a high-level feature representation obtained after convolution, pooling, and activation processing; the category prediction result can be prediction information used to characterize whether a hook-shaped echo exists in the radar echo image to be identified and its category; the spatial positioning result can be used to indicate the spatial location information of the hook-shaped echo in the radar echo image. Based on the third feature map, the third feature map is input to the output layer, and the output layer performs weighted convergence and discrimination operations on the features in the third feature map to generate the category prediction result and the corresponding spatial positioning result of the hook-shaped echo.

[0079] In one embodiment, the output layer may include a classification branch and a localization branch, wherein the classification branch is used to output the category prediction result of the hook echo based on the third feature map, and the localization branch is used to output the spatial localization result corresponding to the hook echo based on the third feature map.

[0080] In one embodiment, the category prediction result may include the classification result of whether the hook echo exists or not, or the target category identifier and corresponding confidence information among multiple echo categories; the spatial positioning result may include the position coordinates, bounding box information or pixel-level region range of the hook echo in the radar echo image, used to describe the spatial distribution of the hook echo.

[0081] In one embodiment, after generating category prediction results and spatial positioning results, the output layer can also perform normalization processing or threshold filtering processing on the prediction results to improve the stability and reliability of the recognition results in practical application scenarios.

[0082] Through the above steps, the category determination and spatial positioning of hook echoes can be achieved simultaneously based on the third feature map, providing accurate and usable identification results for subsequent meteorological analysis, severe convective weather identification, and early warning decision-making.

[0083] S1208. Based on the category prediction result and the spatial positioning result, update the network parameters of the deep convolutional neural network model to obtain the trained deep convolutional neural network model.

[0084] For example, the category prediction result and spatial localization result can be the model prediction output generated by the output layer. The network parameters can be trainable parameters such as the weights of each convolutional kernel, bias parameters, and weights of fully connected layers in the deep convolutional neural network model. The trained deep convolutional neural network model can be a model whose parameters have converged and meet the preset performance indicators in the current training round or training stage. Based on the obtained category prediction result and spatial localization result, the error between the model's predicted output and the corresponding real label result is calculated, and the network parameters of the deep convolutional neural network model are updated based on this error, thereby gradually optimizing the model's ability to identify and locate hook echoes, and obtaining the trained deep convolutional neural network model.

[0085] In one embodiment, the network parameters can be updated by constructing a joint loss function based on the classification loss between the category prediction result and the true category label, and the positioning loss between the spatial positioning result and the true positioning identifier, and then updating the network parameters using the backpropagation algorithm.

[0086] In one embodiment, gradient descent, stochastic gradient descent, or improved algorithms thereof can be used during the model parameter update process to iteratively adjust the network parameters based on the gradient information of the joint loss function until the model training error is less than a preset threshold or the training epochs reach a preset upper limit.

[0087] In one embodiment, when the recognition accuracy, localization accuracy, or loss function trend of the deep convolutional neural network model on the validation set meets the preset conditions, the network parameter update is stopped, and the trained deep convolutional neural network model is output.

[0088] Through the above steps, the deep convolutional neural network model can simultaneously optimize parameters in both category determination and spatial localization dimensions, thereby improving the model's overall recognition performance and practical application effect for hook echoes.

[0089] Optionally, Figure 4 A flowchart of the model training method based on the BCE loss function provided in the embodiments of this application is given. (Reference) Figure 4 The model training method based on the BCE loss function specifically includes:

[0090] S1209. Input the radar echo training image into the deep convolutional neural network model to obtain the predicted probability value.

[0091] For example, the radar echo training image can be a training sample image containing hook-shaped echo identification information, the deep convolutional neural network model can be a model trained through the aforementioned steps and capable of recognizing hook-shaped echoes, and the predicted probability value can be the probability result output by the model for the radar echo training image, representing the likelihood of the existence of hook-shaped echoes. The radar echo training image is input into the deep convolutional neural network model, and through the model's forward inference process, feature extraction, feature fusion, and discrimination operations are performed on the radar echo training image, outputting the corresponding predicted probability value.

[0092] In one embodiment, the predicted probability value can be generated by the probability prediction unit in the output layer to represent the confidence level of the presence of hook echoes in the radar echo training image. The predicted probability value can range from 0 to 1, where a larger value indicates that the model determines that there is a higher probability of the presence of hook echoes.

[0093] In one embodiment, before generating the predicted probability values, the radar echo training images can undergo preprocessing operations consistent with those used in the training phase to ensure that the input data distribution remains consistent with that used in the model training phase, thereby improving the reliability of the predicted probability values.

[0094] In one embodiment, the predicted probability value can serve as the basis for subsequent loss function calculations, model performance evaluations, or threshold determinations, guiding network parameter updates or analyzing the model's recognition performance during the training phase.

[0095] Through the above steps, the predicted probability values ​​output by the deep convolutional neural network model can be obtained based on the radar echo training images, providing a quantitative basis for model training convergence determination, performance evaluation, and subsequent parameter optimization.

[0096] For example, the hook echo identifier can be real label information pre-annotated for radar echo training images, used to indicate whether hook echoes exist in the corresponding training images; the predicted probability value can be the probability of hook echo existence output by the deep convolutional neural network model for the radar echo training images; the binary cross-entropy loss value can be a loss function value used to measure the degree of difference between the model's prediction result and the real identifier. First, the hook echo identifier corresponding to the radar echo training image is obtained; then, the hook echo identifier is compared with the obtained predicted probability value, and the corresponding binary cross-entropy loss value is calculated based on the binary cross-entropy loss function.

[0097] In one embodiment, the binary cross-entropy loss function can be expressed as:

[0098]

[0099] in, Indicates a hook-shaped echo symbol. This represents the predicted probability value. This loss function is used to quantitatively evaluate the deviation between the model's predicted probability and the true label.

[0100] In one embodiment, when the predicted probability value is highly consistent with the hook echo identifier, the corresponding binary cross-entropy loss value is small; when the predicted probability value deviates significantly from the hook echo identifier, the corresponding binary cross-entropy loss value is large, thus reflecting the current prediction performance of the model.

[0101] In one embodiment, the binary cross-entropy loss value can serve as an important basis for updating network parameters, guiding the backpropagation and parameter optimization of the deep convolutional neural network model in subsequent steps.

[0102] Through the above steps, the binary cross-entropy loss value, which reflects the prediction error of the model, can be calculated based on the hook echo identifier and the prediction probability value, providing quantitative support for the training convergence and performance improvement of deep convolutional neural network models.

[0103] S1211. Based on the binary cross-entropy loss value, the parameters of the deep convolutional neural network model are updated by backpropagation to obtain the trained deep convolutional neural network model.

[0104] For example, the binary cross-entropy loss value can be the output value of a loss function used to characterize the degree of error between the model's prediction result and the true identifier of the hook echo. The parameters of the deep convolutional neural network model can include trainable parameters such as the kernel weights and bias parameters in the convolutional layers and the weight parameters in the output layer. The trained deep convolutional neural network model can be a model whose parameters have been optimized and meet the preset performance requirements in the current training stage. Based on the obtained binary cross-entropy loss value, backpropagation is performed on the deep convolutional neural network model, propagating the gradient information corresponding to the loss value from the output layer forward to each network layer, and updating the model parameters according to the gradient information, thereby reducing the model's prediction error.

[0105] In one embodiment, the backpropagation update method for model parameters can be as follows: calculate the gradient corresponding to each network parameter based on the binary cross-entropy loss value, and use gradient descent, stochastic gradient descent or their improved algorithms to iteratively update the model parameters so that the loss value gradually decreases.

[0106] In one embodiment, during the backpropagation update process, a learning rate, regularization parameter, or momentum parameter can be set to control the magnitude of parameter updates and prevent model overfitting or training instability.

[0107] In one embodiment, when the binary cross-entropy loss value of the deep convolutional neural network model drops to the target threshold within a preset training round, or when the recognition performance on the validation set meets the preset conditions, the backpropagation update is stopped, and the trained deep convolutional neural network model is output.

[0108] Through the above steps, the parameters of the deep convolutional neural network model can be effectively optimized based on the binary cross-entropy loss value, which significantly improves the prediction accuracy and stability of the model in the hook echo recognition task, thereby obtaining a trained deep convolutional neural network model.

[0109] Optionally, Figure 5 A flowchart of the model training method based on the DFL loss function provided in the embodiments of this application is given. (Reference) Figure 5 The model training method based on the DFL loss function specifically includes:

[0110] S1212. Input the radar echo training image into the deep convolutional neural network model to obtain the bounding box regression prediction result.

[0111] For example, the radar echo training image can be a training sample image containing hook-shaped echo identification information, the deep convolutional neural network model can be a model trained through the aforementioned steps and capable of hook-shaped echo recognition and localization, and the bounding box regression prediction result can be the regression result output by the model for the radar echo training image, used to characterize the spatial range of the hook-shaped echo. The radar echo training image is input into the deep convolutional neural network model. Through the model's forward inference process, feature extraction and spatial relationship modeling are performed on the radar echo training image, and the corresponding bounding box regression prediction result is output by the regression branch of the model.

[0112] In one embodiment, the bounding box regression prediction result may include multiple regression parameters for describing the spatial location of the hook echo, such as the center coordinates, width, and height of the bounding box, or the coordinates of the upper left and lower right corners of the bounding box, to accurately characterize the spatial distribution range of the hook echo in the radar echo training image.

[0113] In one embodiment, the bounding box regression prediction result can be generated based on the third feature map or higher-level semantic features, so that the regression result simultaneously integrates the morphological features of the hook echo and spatial context information, thereby improving the accuracy of spatial positioning.

[0114] In one embodiment, the bounding box regression prediction result can be used as the basis for subsequent regression loss calculation and network parameter update, and can be used to further optimize the deep convolutional neural network model's ability to spatially locate hook echoes during the training phase.

[0115] Through the above steps, the bounding box regression prediction results output by the deep convolutional neural network model can be obtained based on the radar echo training images, providing important input for the accurate spatial localization of hook echoes and subsequent model parameter optimization.

[0116] S1213. Obtain the labeled bounding box of the radar echo training image, and calculate the distribution focus loss value based on the labeled bounding box and the bounding box regression prediction result.

[0117] For example, the labeled bounding box can be the pre-annotated real spatial positioning information of the hook-shaped echo region in the radar echo training image, used to characterize the actual location range of the hook-shaped echo in the image; the bounding box regression prediction result can be the predicted bounding box parameters output by the deep convolutional neural network model for the radar echo training image; the distributed focus loss value can be the regression loss function value used to measure the distribution difference between the predicted bounding box and the labeled bounding box. First, the labeled bounding box corresponding to the radar echo training image is obtained; then, the labeled bounding box is compared and analyzed with the obtained bounding box regression prediction result, and the corresponding distributed focus loss value is calculated based on the distributed focus loss function.

[0118] In one embodiment, the distributed focus loss function can finely characterize the bounding box regression error by modeling the bounding box position parameters as a discrete probability distribution and calculating the difference between the predicted distribution and the true distribution, thus making the model more focused on the regression accuracy of the boundary position.

[0119] In one embodiment, when the bounding box regression prediction result is close to the labeled bounding box in terms of spatial location and scale distribution, the corresponding distribution focus loss value is small; when the difference between the two is large, the corresponding distribution focus loss value is large, thus reflecting the accuracy of the model's current spatial positioning prediction.

[0120] In one embodiment, the distributed focus loss value can serve as an important basis for subsequent network parameter updates, guiding the deep convolutional neural network model to optimize the parameters of the bounding box regression branch during backpropagation.

[0121] Through the above steps, the distribution focal loss value reflecting the spatial positioning error can be calculated based on the identified bounding box and the bounding box regression prediction results, providing an effective loss constraint for improving the positioning accuracy of the deep convolutional neural network model for hook echo boundaries.

[0122] S1214. Based on the distributed focus loss value, the parameters of the deep convolutional neural network model are updated by backpropagation to obtain the trained deep convolutional neural network model.

[0123] For example, the distributed focus loss value can be the output value of a loss function used to characterize the spatial distribution difference between the bounding box regression prediction result and the identified bounding box. The parameters of the deep convolutional neural network model can include the weight parameters in the regression branch, the convolution kernel parameters in the convolutional layer, and the weight parameters in the shared feature layer. The trained deep convolutional neural network model can be a model whose bounding box regression accuracy reaches the preset requirement in the current training stage. Based on the obtained distributed focus loss value, backpropagation is performed on the deep convolutional neural network model, propagating the gradient information corresponding to the distributed focus loss value from the regression output layer forward to each network layer, and updating the model parameters according to the gradient information, thereby reducing the bounding box regression prediction error.

[0124] In one embodiment, the backpropagation update method for model parameters can be as follows: calculate the gradient corresponding to each network parameter based on the distributed focus loss value, and use gradient descent, stochastic gradient descent or their improved algorithms to iteratively update the model parameters so that the distributed focus loss value gradually decreases.

[0125] In one embodiment, different learning rates or weight coefficients can be set for the regression branch and the classification branch during the parameter update process to balance the optimization effects of the two tasks, category recognition and spatial localization, during the model training process.

[0126] In one embodiment, when the bounding box localization error or distribution focus loss value of the deep convolutional neural network model on the validation set meets the preset convergence condition, backpropagation updates are stopped, and the trained deep convolutional neural network model is output.

[0127] Through the above steps, the parameters of the deep convolutional neural network model can be optimized in a targeted manner based on the distributed focus loss value, so as to further improve the prediction accuracy and stability of the model in the hook echo spatial localization task, thereby obtaining the trained deep convolutional neural network model.

[0128] Optionally, Figure 6 A flowchart of the model training method based on the CIoU loss function provided in this application is given. (Reference) Figure 6 The model training method based on the CIoU loss function specifically includes:

[0129] S1215. Input the radar echo training image into the deep convolutional neural network model to obtain the predicted bounding box.

[0130] For example, the radar echo training image can be a training sample image containing hook-shaped echo identification information, the deep convolutional neural network model can be a model that has been jointly trained and optimized using the aforementioned classification loss and regression loss, and the predicted bounding box can be the model's output prediction result for the radar echo training image, used to characterize the spatial location range of the hook-shaped echo. The radar echo training image is input into the deep convolutional neural network model. Through the model's forward inference process, feature extraction, feature fusion, and spatial relationship modeling are performed on the radar echo training image, and the model's bounding box prediction branch outputs the corresponding predicted bounding box.

[0131] In one embodiment, the predicted bounding box can be represented in a parametric form, including the center coordinates, width, and height of the bounding box, or the coordinates of the top-left and bottom-right corners of the bounding box, to describe the spatial coverage of the hook echo in the radar echo training image.

[0132] In one embodiment, the predicted bounding box can be generated based on a third feature map or higher-level semantic features, so that the prediction result simultaneously integrates the morphological features of the hook echo, spatial context information, and multi-scale feature information, thereby improving the accuracy of the bounding box prediction.

[0133] In one embodiment, the predicted bounding box can be used as input data for subsequent localization accuracy evaluation, bounding box consistency verification, or model performance verification to test the actual performance of the deep convolutional neural network model in the hook echo spatial localization task.

[0134] Through the above steps, the predicted bounding boxes of the deep convolutional neural network model output can be obtained based on radar echo training images during the model training or validation phase, providing a direct basis for the accurate localization of hook echoes and the evaluation of model performance.

[0135] S1216. Obtain the labeled bounding box of the radar echo training image, and calculate the CIoU loss value based on the predicted bounding box and the labeled bounding box.

[0136] For example, the labeled bounding box can be the real spatial positioning information pre-annotated for the hook-shaped echo region in the radar echo training image; the predicted bounding box can be the bounding box prediction result output by the deep convolutional neural network model in step S1215; the CIoU loss value can be a regression loss function value used to comprehensively measure the overlap, center point distance, and aspect ratio consistency between the predicted bounding box and the labeled bounding box. First, the labeled bounding box corresponding to the radar echo training image is obtained; then, the predicted bounding box and the labeled bounding box are aligned and compared, and the corresponding CIoU loss value is calculated based on the CIoU (Complete Intersection over Union) loss function.

[0137] In one embodiment, the CIoU loss function not only considers the intersection-union ratio (IoU) between the predicted bounding box and the labeled bounding box during the calculation process, but also introduces the distance constraint between the center points of the predicted bounding box and the labeled bounding box, as well as the aspect ratio difference constraint between the two, so as to provide a more comprehensive and detailed characterization of the bounding box regression error.

[0138] In one embodiment, when the predicted bounding box and the labeled bounding box are highly consistent in spatial location, scale, and shape ratio, the corresponding CIoU loss value is small; when there are large differences between the two in terms of positional offset, scale, or shape, the corresponding CIoU loss value is large, thus reflecting the accuracy of the model's current bounding box prediction results.

[0139] In one embodiment, the CIoU loss value can serve as an important basis for subsequent backpropagation updates of network parameters, guiding the deep convolutional neural network model to further optimize the parameters of the bounding box regression branch.

[0140] Through the above steps, the CIoU loss value, which reflects the combined error of spatial location, scale, and shape, can be calculated based on the predicted bounding box and the labeled bounding box, providing an effective loss constraint for improving the overall accuracy of deep convolutional neural network models in locating hook echo boundaries.

[0141] S1217. Based on the CIoU loss value, backpropagation is performed to update the parameters of the deep convolutional neural network model to obtain the trained deep convolutional neural network model.

[0142] For example, the CIoU loss value can be the output value of a regression loss function used to comprehensively characterize the differences between the predicted bounding box and the labeled bounding box in terms of overlap, center distance, and aspect ratio consistency. The parameters of the deep convolutional neural network model can include the weight parameters in the bounding box regression branch, the convolution kernel parameters in the shared feature extraction layer, and the relevant parameters in the output layer. The trained deep convolutional neural network model can be a model whose bounding box localization performance meets the preset requirements in the current training stage. Based on the obtained CIoU loss value, backpropagation is performed on the deep convolutional neural network model, propagating the gradient information corresponding to the CIoU loss value from the bounding box regression output layer forward to each network layer of the model, and updating the model parameters according to the gradient information, thereby reducing the comprehensive localization error between the predicted bounding box and the labeled bounding box.

[0143] In one embodiment, the backpropagation update method for model parameters can be as follows: calculate the gradient corresponding to each network parameter based on the CIoU loss value, and use gradient descent, stochastic gradient descent or their improved algorithms to iteratively update the model parameters so that the CIoU loss value gradually decreases.

[0144] In one embodiment, during the parameter update process, an independent learning rate or loss weight coefficient can be set for the bounding box regression branch to enhance the model's ability to optimize spatial localization accuracy, while avoiding adverse effects on the training effect of the classification branch.

[0145] In one embodiment, when the CIoU loss value of the deep convolutional neural network model on the validation dataset drops to a preset threshold, or the bounding box localization accuracy reaches a preset target, the backpropagation update is stopped, and the trained deep convolutional neural network model is output.

[0146] Through the above steps, the parameters of the deep convolutional neural network model can be optimized based on the CIoU loss value, which further improves the spatial accuracy and overall stability of the model in the hook echo boundary localization task, thereby obtaining a trained deep convolutional neural network model.

[0147] Optionally, Figure 7 A flowchart of the model verification method provided in this application is given. (Reference) Figure 7 The model validation method specifically includes:

[0148] S121. Generate a radar echo test image based on the historical meteorological radar echo image.

[0149] For example, historical weather radar echo images can be raw echo image data collected and stored by weather radar at different times and under different weather conditions, while radar echo test images can be a set of sample images selected or constructed from historical weather radar echo images for model testing and performance evaluation. Generating radar echo test images from historical weather radar echo images can specifically include selecting echo images from historical weather radar echo images that were not used in model training as test samples, or generating radar echo test images for testing after pre-processing historical weather radar echo images.

[0150] In one embodiment, the radar echo test image can be generated by extracting corresponding image data from historical meteorological radar echo images according to a preset time interval, radar station, or weather type to construct a representative test sample set.

[0151] In one embodiment, during the generation of radar echo test images, the test images may undergo preprocessing operations consistent with those in the training phase, including image resizing, pixel value normalization, or format conversion, to ensure consistency between the radar echo test images and the input requirements of the deep convolutional neural network model.

[0152] In one embodiment, the radar echo test image may include test samples containing hook-shaped echoes and control samples not containing hook-shaped echoes, for comprehensively evaluating the recognition performance and generalization ability of the deep convolutional neural network model in different scenarios.

[0153] Through the above steps, a radar echo test image set independent of the training data can be constructed based on historical meteorological radar echo images, providing a reliable data foundation for the subsequent testing, verification, and performance evaluation of the hook echo recognition model.

[0154] S122. Input the radar echo test image into the trained deep convolutional neural network model to obtain the hook echo recognition test result.

[0155] For example, the radar echo test image can be a test sample image constructed based on historical meteorological radar echo images and not involved in the model training process. The trained deep convolutional neural network model can be a model that has undergone parameter optimization through the aforementioned training steps and possesses the ability to identify and locate hook echoes. The hook echo recognition test result can be the recognition and location result output by the model for the radar echo test image. The radar echo test image is input into the trained deep convolutional neural network model. Through the model's forward inference process, feature extraction and discriminant analysis are performed on the radar echo test image to obtain the corresponding hook echo recognition test result.

[0156] In one embodiment, the hook echo recognition test result may include the result of determining the presence or absence of the hook echo, the category prediction result, and the corresponding prediction probability value, which are used to reflect the confidence level of the model in recognizing the hook echo in the test image.

[0157] In one embodiment, the hook echo recognition test results may also include the spatial localization results of the hook echo in the radar echo test image, such as predicted bounding boxes, region contours, or pixel-level localization information, to evaluate the accuracy of the model in spatial localization.

[0158] In one embodiment, after obtaining the hook echo recognition test results, the recognition test results can be further compared and analyzed with the pre-annotated test markers in the radar echo test image to evaluate the recognition accuracy, positioning accuracy and generalization performance of the trained deep convolutional neural network model on the test dataset.

[0159] Through the above steps, the trained deep convolutional neural network model can be effectively tested based on independent radar echo test images, thereby verifying the model's practical application effect and reliability in the hook echo recognition task.

[0160] S123. Obtain the identification information of the radar echo test image, and calculate the precision, recall and F1 score of the trained deep convolutional neural network model based on the hook echo recognition test results and the identification information.

[0161] For example, the identification information of the radar echo test image can be the real results of pre-annotating the presence of hook-shaped echoes and their corresponding positions in the test image, serving as a benchmark for the model testing phase; the hook-shaped echo recognition test result can be the recognition result output by the trained deep convolutional neural network model; precision, recall, and F1 score can be evaluation metrics used to measure the model's recognition performance. First, the identification information corresponding to the radar echo test image is obtained; then, the identification information is compared and analyzed with the hook-shaped echo recognition test result, and the precision, recall, and F1 score of the trained deep convolutional neural network model are calculated based on the comparison results.

[0162] In one embodiment, precision is used to represent the proportion of samples that the model predicts as hook echoes but are actually hook echoes, recall is used to represent the proportion of samples in the test image that actually have hook echoes that are correctly identified by the model, and F1 score is used to comprehensively reflect the balance between precision and recall.

[0163] In one embodiment, precision, recall, and F1 score can be calculated based on a confusion matrix formed between the model recognition results and the identification information, wherein the confusion matrix includes statistical results of true positive, false positive, true negative, and false negative samples.

[0164] In one embodiment, by calculating the precision, recall, and F1 score of the trained deep convolutional neural network model on radar echo test images, the model's recognition accuracy, missed detection rate, and overall performance in the hook echo recognition task can be comprehensively evaluated.

[0165] The above steps enable a quantitative evaluation of the recognition performance of the trained deep convolutional neural network model, providing an objective performance reference for model optimization, parameter adjustment, and practical application deployment.

[0166] S130. Obtain the echo image to be identified, and use the trained deep convolutional neural network model to identify the hook-shaped echo in the echo image to be identified.

[0167] In some embodiments, the echo image to be identified can be radar echo image data acquired by a weather radar at the current time or during the target monitoring period. The trained deep convolutional neural network model can be a model trained based on historical radar echo training images and capable of identifying hook echoes. Hook echo identification is a process of determining whether hook echoes exist in the radar echo image and their corresponding locations. First, the echo image to be identified is acquired; then, the echo image to be identified is input into the trained deep convolutional neural network model, which performs feature extraction and pattern matching on the echo image to be identified, thereby completing the identification of hook echoes.

[0168] In one embodiment, the method of using a trained deep convolutional neural network model to identify hook echoes can be as follows: through the forward inference process of the model, the identification result representing the probability of the existence of hook echoes, category labels, or corresponding spatial regions is output.

[0169] In one embodiment, the hook echo recognition result may include hook echo recognition confidence information, pixel region or boundary information corresponding to the hook echo, which are used to indicate the specific location and morphological features of the hook echo in the echo image to be recognized.

[0170] In one embodiment, after completing the hook echo identification, post-processing operations can be performed on the identification results, such as filtering the identification results based on a confidence threshold, merging or smoothing adjacent identification regions, to improve the accuracy and stability of the hook echo identification results.

[0171] Optionally, the hook echo recognition method provided in this application embodiment is based on a supervised learning deep convolutional neural network to achieve automatic recognition and spatial positioning of hook echoes from meteorological radar.

[0172] In one embodiment, strong convective echo images from S-band dual-polarization weather radar in Guangzhou over the past six years were collected as the original data source. For radar image samples containing hook echoes, target areas were manually labeled to construct a high-quality labeled dataset. Considering the multi-elevation angle scanning capability of the Guangzhou S-band dual-polarization radar, a single radar detection can acquire multiple hook echo samples at different elevation angles, thereby improving the spatial diversity of the samples. To alleviate the problem of insufficient hook echo samples, data augmentation processing was performed on image samples containing hook echoes, specifically by rotational augmentation with a step size of 10° to generate multi-angle positive samples; radar images without hook echoes were used as negative samples. The ratio of positive to negative samples was controlled at approximately 1:3, with 760 positive samples and 2456 negative samples. All samples were randomly divided into training, validation, and test sets in a ratio of 7:2:1 to ensure the reliability of model training and evaluation.

[0173] In one embodiment, a lightweight Generalized Efficient Layer Aggregation Network (GELAN) is introduced as the backbone network in terms of model architecture design to balance inference speed and detection accuracy. A deep convolutional neural network model is constructed based on the PyTorch deep learning framework. The model consists of an input layer, a feature extraction hidden layer, and an output layer. The hidden layer comprises multiple levels of convolutional layers, pooling layers, and activation layers. The convolutional layers extract spatial features from the radar echo image, the pooling layers enhance the robustness of the features through dimensionality reduction, and the activation layers use the SiLU activation function to enhance nonlinear expressiveness and effectively alleviate the gradient vanishing problem. The input layer receives radar echo image information, and the output layer simultaneously performs hook echo category determination and target location regression, achieving integrated recognition and localization processing.

[0174] In one embodiment, during the model training phase, a multi-task joint loss function is constructed to collaboratively optimize the classification and localization processes. For target category determination, a Binary Cross Entropy (BCE) loss function is used to measure the accuracy of hook echo category prediction. For target location regression, a Distribution Focal Loss (DFL) and a CIoU bounding box loss function are introduced. DFL improves the granularity of bounding box regression by modeling the probability distribution of bounding box coordinates; CIoU loss comprehensively considers the overlap area between the predicted and ground truth boxes, the distance between their center points, and the aspect ratio difference to more accurately constrain target localization deviation. During backpropagation, an SGD optimizer is used to iteratively update the network parameters, causing the overall loss function to gradually converge, thereby obtaining a deep convolutional neural network model with good fitting ability and generalization performance.

[0175] In one embodiment, the trained model is validated using a test set to assess its generalization ability on unknown samples. Precision, recall, and F1 score are used as comprehensive evaluation metrics to quantitatively analyze the hook echo recognition performance. Experimental results show that, with the SGD optimizer, the model achieves a precision of 0.946, a recall of 0.913, and an F1 score of 0.930 on the test set, indicating that the model has high stability and reliability in hook echo recognition and localization tasks.

[0176] In one embodiment, a hook echo recognition and localization visualization product for S-band dual-polarization radar is constructed based on a trained deep convolutional neural network model. In practical applications, the model automatically analyzes the input radar echo image and marks the spatial location of the hook echoes with rectangular bounding boxes, providing an intuitive display of the recognition results. Experiments show that this method achieves a precision, recall, and F1 score exceeding 0.9 for hook echo recognition on S-band dual-polarization radar images, meeting the needs of rapid identification and accurate localization of severe convective weather features in operational applications.

[0177] Based on the above embodiments, Figure 8 This is a schematic diagram of the hook-shaped echo recognition device provided in an embodiment of this application. (Reference) Figure 8 The hook-shaped echo recognition device provided in this embodiment specifically includes: a data identification module 21, a model training module 22, and an echo recognition module 23.

[0178] The data identification module 21 is configured to acquire historical meteorological radar echo images, identify hook-shaped echoes in the historical meteorological radar echo images, and obtain radar echo training images; the model training module 22 is configured to train a deep convolutional neural network model based on the radar echo training images to obtain a trained deep convolutional neural network model; and the echo recognition module 23 is configured to acquire an echo image to be recognized, and use the trained deep convolutional neural network model to recognize hook-shaped echoes in the echo image to be recognized.

[0179] Based on the above embodiments, the radar echo training image includes a first training image with hook-shaped echoes and a second training image without hook-shaped echoes. The model training module 22 includes: an image rotation unit configured to rotate the first training image at preset angular intervals to obtain a third training image; and a first training unit configured to train the deep convolutional neural network model using the first training image, the second training image, and the third training image to obtain a trained deep convolutional neural network model.

[0180] Based on the above embodiments, the model training module 22 includes: an input layer unit configured to input the radar echo training image into the input layer of the deep convolutional neural network model; a first feature unit configured to extract features from the radar echo training image through the convolutional layer of the deep convolutional neural network model to obtain a first feature map; a second feature unit configured to perform pooling processing on the first feature map to obtain a second feature map; a third feature unit configured to perform activation processing on the second feature map using the SiLU activation function to obtain a third feature map; an output layer unit configured to generate a category prediction result and a spatial positioning result of the hook echo through the output layer based on the third feature map; and a second training unit configured to update the network parameters of the deep convolutional neural network model according to the category prediction result and the spatial positioning result to obtain a trained deep convolutional neural network model.

[0181] Based on the above embodiments, the model training module 22 includes: a prediction probability unit, configured to input the radar echo training image into a deep convolutional neural network model to obtain a prediction probability value; a first loss unit, configured to acquire the hook echo identifier of the radar echo training image and calculate the binary cross-entropy loss value between the hook echo identifier and the prediction probability value; and a third training unit, configured to perform backpropagation update on the parameters of the deep convolutional neural network model based on the binary cross-entropy loss value to obtain a trained deep convolutional neural network model.

[0182] Based on the above embodiments, the model training module 22 includes: a regression prediction unit configured to input the radar echo training image into a deep convolutional neural network model to obtain a bounding box regression prediction result; a second loss unit configured to obtain the labeled bounding box of the radar echo training image and calculate a distribution focus loss value based on the labeled bounding box and the bounding box regression prediction result; and a fourth training unit configured to perform backpropagation update on the parameters of the deep convolutional neural network model based on the distribution focus loss value to obtain a trained deep convolutional neural network model.

[0183] Based on the above embodiments, the model training module 22 includes: a bounding box prediction unit, configured to input the radar echo training image into a deep convolutional neural network model to obtain predicted bounding boxes; a third loss unit, configured to obtain the labeled bounding boxes of the radar echo training image and calculate CIoU loss values ​​based on the predicted bounding boxes and the labeled bounding boxes; and a fifth training unit, configured to perform backpropagation updates on the parameters of the deep convolutional neural network model based on the CIoU loss values ​​to obtain a trained deep convolutional neural network model.

[0184] Based on the above embodiments, the hook echo recognition device further includes: a test image module configured to generate a radar echo test image based on the historical meteorological radar echo image; a test input module configured to input the radar echo test image into a trained deep convolutional neural network model to obtain a hook echo recognition test result; and a test evaluation module configured to acquire the identification information of the radar echo test image and calculate the precision, recall, and F1 score of the trained deep convolutional neural network model based on the hook echo recognition test result and the identification information.

[0185] The hook echo recognition device provided in this application embodiment, through the coordinated operation of its various functional modules, achieves a complete closed-loop processing flow from historical sample construction and model training to actual echo recognition, effectively improving the automation level and recognition accuracy of hook echo recognition. Specifically, the data labeling module 21 is configured to acquire historical meteorological radar echo images and perform manual or semi-automatic labeling of hook echo targets in these images to generate radar echo training images for model training, thereby providing a training sample basis with clear semantic labels for subsequent model learning. The model training module 22 works in conjunction with the data labeling module 21 and is configured to iteratively train a preset deep convolutional neural network model based on the radar echo training images. Through feature extraction, parameter updates, and loss function optimization, the model's ability to represent and distinguish hook echo features is gradually improved, ultimately resulting in a trained deep convolutional neural network model. The echo recognition module 23 is configured to acquire the meteorological radar echo image to be identified and call the trained deep convolutional neural network model to perform inference analysis on the echo image to be identified, thereby realizing the automatic identification of hook-shaped echo targets and outputting the corresponding recognition results for subsequent display, early warning or business decision-making. Through the streamlined collaboration of the above modules, this embodiment realizes a full-process processing mechanism from historical radar echo data identification and model training to real-time echo recognition, which not only ensures the accuracy and consistency of hook-shaped echo recognition results, but also improves the processing efficiency and application stability of the system in actual meteorological business scenarios, and is suitable for intelligent recognition applications of meteorological radar hook-shaped echoes in multiple time periods and multiple regions.

[0186] The hook-shaped echo recognition device provided in this application embodiment can be used to execute the hook-shaped echo recognition method provided in the above embodiment, and has corresponding functions and beneficial effects.

[0187] Figure 9 This is a schematic diagram of the structure of a hook-shaped echo identification device provided in an embodiment of this application, with reference to... Figure 9The hook-shaped echo identification device includes a processor 31, a memory 32, a communication device 33, an input device 34, and an output device 35. The number of processors 31 and the number of memories 32 in the hook-shaped echo identification device can be one or more. The processor 31, memory 32, communication device 33, input device 34, and output device 35 of the hook-shaped echo identification device can be connected via a bus or other means.

[0188] The memory 32, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as program instructions / modules corresponding to the hook-echo recognition method in any embodiment of this application (e.g., data identification module 21, model training module 22, and echo recognition module 23 in the hook-echo recognition device). The memory 32 may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created based on the use of the device, etc. Furthermore, the memory 32 may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the device via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0189] The communication device 33 is used for data transmission.

[0190] The processor 31 executes various functional applications and data processing of the device by running software programs, instructions and modules stored in the memory 32, thereby realizing the hook echo recognition method described above.

[0191] Input device 34 can be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the device. Output device 35 may include display devices such as a display screen.

[0192] The hook-shaped echo recognition device provided above can be used to execute the hook-shaped echo recognition method provided in the above embodiments, and has corresponding functions and beneficial effects.

[0193] This application embodiment also provides a storage medium containing computer-executable instructions. When executed by a computer processor, the computer-executable instructions are used to perform a hook echo recognition method. The hook echo recognition method includes: acquiring historical weather radar echo images; identifying hook echoes in the historical weather radar echo images to obtain radar echo training images; training a deep convolutional neural network model based on the radar echo training images to obtain a trained deep convolutional neural network model; acquiring an echo image to be recognized; and using the trained deep convolutional neural network model to perform hook echo recognition on the echo image to be recognized.

[0194] Storage medium—any type of memory device or storage device. The term "storage medium" is intended to include: mounting media, such as CD-ROM, floppy disk, or magnetic tape devices; computer system memory or random access memory, such as DRAM, DDR RAM, SRAM, EDO RAM, etc.; non-volatile memory, such as flash memory, magnetic media (e.g., hard disk or optical storage); registers or other similar types of memory elements, etc. Storage medium may also include other types of memory or combinations thereof. Furthermore, storage medium may reside in a first computer system in which a program is executed, or it may reside in a different second computer system connected to the first computer system via a network (such as the Internet). The second computer system can provide program instructions to the first computer for execution. The term "storage medium" may include two or more storage media residing in different locations (e.g., in different computer systems connected via a network). Storage medium may store program instructions (e.g., specifically implemented as a computer program) executable by one or more processors.

[0195] Of course, the computer-executable instructions provided in the embodiments of this application are not limited to the hook echo recognition method described above, but can also perform related operations in the hook echo recognition method provided in any embodiment of this application.

[0196] The hook echo identification device, storage medium, and hook echo identification equipment provided in the above embodiments can execute the hook echo identification method provided in any embodiment of this application. For technical details not described in detail in the above embodiments, please refer to the hook echo identification method provided in any embodiment of this application.

[0197] The above description is merely a preferred embodiment and the technical principles employed in this application. This application is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions that can be made by those skilled in the art will not depart from the scope of protection of this application. Therefore, although this application has been described in detail through the above embodiments, this application is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of this application. The scope of this application is determined by the scope of the claims.

Claims

1. A hook-shaped echo recognition method, characterized in that, include: Acquire historical weather radar echo images, identify hook-shaped echoes in the historical weather radar echo images, and obtain radar echo training images. The deep convolutional neural network model is trained based on the radar echo training image to obtain the trained deep convolutional neural network model. Obtain the echo image to be identified, and use the trained deep convolutional neural network model to identify the hook-shaped echo in the echo image.

2. The hook-shaped echo recognition method according to claim 1, characterized in that, The radar echo training images include a first training image with hook-shaped echoes and a second training image without hook-shaped echoes. The training of the deep convolutional neural network model based on the radar echo training images to obtain the trained deep convolutional neural network model includes: The first training image is rotated at preset angular intervals to obtain the third training image; The deep convolutional neural network model is trained using the first training image, the second training image, and the third training image to obtain a trained deep convolutional neural network model.

3. The hook-shaped echo recognition method according to claim 1, characterized in that, The step of training the deep convolutional neural network model based on the radar echo training image to obtain the trained deep convolutional neural network model includes: The radar echo training image is input into the input layer of the deep convolutional neural network model; The radar echo training image is used to extract features through the convolutional layer of the deep convolutional neural network model to obtain a first feature map; The first feature map is subjected to pooling to obtain the second feature map; The second feature map is activated using the SiLU activation function to obtain the third feature map; Based on the third feature map, the hook echo category prediction result and spatial localization result are generated through the output layer; Based on the category prediction results and the spatial positioning results, the network parameters of the deep convolutional neural network model are updated to obtain the trained deep convolutional neural network model.

4. The hook-shaped echo recognition method according to claim 1, characterized in that, The step of training the deep convolutional neural network model based on the radar echo training image to obtain the trained deep convolutional neural network model includes: The radar echo training image is input into a deep convolutional neural network model to obtain the predicted probability value; Obtain the hook-shaped echo identifier of the radar echo training image, and calculate the binary cross-entropy loss value between the hook-shaped echo identifier and the predicted probability value. The parameters of the deep convolutional neural network model are updated by backpropagation based on the binary cross-entropy loss value, resulting in a trained deep convolutional neural network model.

5. The hook-shaped echo recognition method according to claim 1, characterized in that, The step of training the deep convolutional neural network model based on the radar echo training image to obtain the trained deep convolutional neural network model includes: The radar echo training image is input into a deep convolutional neural network model to obtain the bounding box regression prediction result; Obtain the labeled bounding box of the radar echo training image, and calculate the distribution focus loss value based on the labeled bounding box and the bounding box regression prediction result; The parameters of the deep convolutional neural network model are updated by backpropagation based on the distributed focus loss value, resulting in a trained deep convolutional neural network model.

6. The hook-shaped echo recognition method according to claim 1, characterized in that, The step of training the deep convolutional neural network model based on the radar echo training image to obtain the trained deep convolutional neural network model includes: The radar echo training image is input into a deep convolutional neural network model to obtain the predicted bounding box; Obtain the labeled bounding box of the radar echo training image, and calculate the CIoU loss value based on the predicted bounding box and the labeled bounding box; The parameters of the deep convolutional neural network model are updated by backpropagation based on the CIoU loss value, resulting in a trained deep convolutional neural network model.

7. The hook-shaped echo recognition method according to claim 1, characterized in that, After training the deep convolutional neural network model based on the radar echo training image to obtain the trained deep convolutional neural network model, the method further includes: A radar echo test image is generated based on the historical meteorological radar echo image; The radar echo test image is input into the trained deep convolutional neural network model to obtain the hook echo recognition test result; Obtain the identification information of the radar echo test image, and calculate the precision, recall and F1 score of the trained deep convolutional neural network model based on the hook echo recognition test results and the identification information.

8. A hook-shaped echo identification device, characterized in that, include: The data identification module is used to acquire historical weather radar echo images, identify the hook-shaped echoes in the historical weather radar echo images, and obtain radar echo training images. The model training module is used to train the deep convolutional neural network model based on the radar echo training image to obtain the trained deep convolutional neural network model. The echo recognition module is used to acquire the echo image to be recognized and to perform hook echo recognition on the echo image to be recognized using the trained deep convolutional neural network model.

9. A hook-shaped echo identification device, characterized in that, include: One or more processors; A memory that stores one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the hook echo recognition method as described in any one of claims 1-7.

10. A storage medium containing computer-executable instructions, characterized in that, The computer-executable instructions, when executed by a computer processor, are used to perform the hook-echo recognition method as described in any one of claims 1-7.

Images