Weather system identification method based on multi-source meteorological information

Abstract

The invention discloses a weather system identification method based on multi-source meteorological information. The method comprises the following steps: S1, acquiring historical multi-source meteorological data of a target area; s2, cleaning and complementing the original meteorological data; s3, marking the type and position of a weather system, and constructing a training set; s4, training a single-source meteorological feature extraction model; s5, fusing the single-source features, and training an integrated recognition model; s6, deploying the trained model to a reasoning environment; and S7, inputting real-time meteorological data, and outputting an identification result. According to the invention, efficient fusion and accurate identification of multi-source weather information are realized, and the real-time performance and accuracy of weather system identification are significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of weather system identification, and in particular to a weather system identification method based on multi-source meteorological information. Background Art

[0002] Weather systems, including convective systems, frontal systems, tropical cyclones, and other types, are key drivers of regional weather variability and the generation and propagation of extreme weather events. Accurately identifying the spatial structure, evolutionary trends, and types of weather systems is fundamental to achieving high-quality weather forecasts, early disaster warnings, and risk-based decision support. Currently, mainstream weather system identification methods primarily include those based on manual interpretation and algorithmic identification methods based on numerical simulation parameters.

[0003] Manual interpretation relies primarily on experienced professional meteorologists to comprehensively analyze and interpret multi-source meteorological imagery, including satellite cloud images, radar echo maps, and pressure field maps. This approach, combined with meteorological expertise, allows for subjective analysis and weather system location. While this approach offers some flexibility, it relies heavily on individual judgment, lacks objectivity and standardization, and is unable to adapt to the high-frequency dynamic recognition requirements of complex weather scenarios. Furthermore, faced with the massive volume of real-time meteorological data, this approach faces significant bottlenecks in processing efficiency and labor costs, making it difficult to support the demands of modern automated business systems.

[0004] Another approach relies on the meteorological element fields output by numerical weather forecast models, performing pattern recognition by setting physical quantity thresholds, gradient change rates, and geopotential height structures. While this approach offers a degree of objectivity and batch processing capabilities, its effectiveness relies on the accuracy and spatial resolution of the numerical model. This is particularly true when dealing with sudden severe convective weather events at the local scale, often leading to problems such as forecast offsets, structural ambiguity, and type recognition errors. Furthermore, due to the complexity of the numerical models and the large amount of computation required, their deployment and rapid inference in real-time systems present significant obstacles, making it difficult to meet the practical needs of rapid response and regionally customized recognition.

[0005] Amid the rapid development of artificial intelligence (AI) technology, researchers have begun applying deep learning models to weather image recognition and system identification tasks. Some work has used convolutional neural networks (CNNs) to train radar imagery, cloud maps, and other applications for precipitation forecasting and thunderstorm identification. However, most methods still rely on a single meteorological data set as input and fail to fully integrate multi-source meteorological information. Furthermore, existing methods often employ an end-to-end, single-unit network architecture, which fails to explicitly reflect the independent contributions of different meteorological elements to the recognition task. Feature fusion can be redundant or conflicting, limiting further improvements in recognition accuracy.

[0006] In summary, the existing technologies still have much room for improvement in terms of multi-source meteorological information fusion, weather system spatial positioning and identification accuracy, which is specifically manifested in the following aspects: First, there is a lack of a unified architecture to access, preprocess and organize data from multiple observation systems, resulting in low information utilization; second, existing deep models mostly use mixed input to perform overall modeling, ignoring the differences in spatial structure and physical meaning of different meteorological elements, and it is difficult to give full play to the advantages of single-source data; third, there is still a lack of a hierarchical neural network system that can construct "first extract features by source, then integrate and identify" to improve the structural clarity and output interpretability of the overall identification; fourth, the existing model deployment mechanism is generally disconnected from the business system, and fails to output structured, screened and optimized identification results in combination with actual meteorological scenarios, making it difficult to directly apply them to early warning platforms and visualization systems.

[0007] Therefore, how to provide a weather system identification method based on multi-source meteorological information is a problem that technicians in this field urgently need to solve. Summary of the Invention

[0008] One purpose of the present invention is to propose a weather system identification method based on multi-source meteorological information. The present invention makes full use of multi-source heterogeneous information such as wind speed, humidity, air pressure, satellite cloud images and radar images in meteorological big data, constructs a deep neural network structure that combines single-source feature extraction with integrated recognition fusion, and describes in detail the algorithm process of realizing intelligent identification of weather system type and location through reconstructing the Backbone-Neck-Head module through the YOLOv5 network. The method has the advantages of strong data adaptability, high recognition accuracy, fast inference speed and easy deployment and application.

[0009] A weather system identification method based on multi-source meteorological information according to an embodiment of the present invention includes the following steps:

[0010] S1. Obtain meteorological data from multiple sources to form an original data set;

[0011] S2. Performing data cleaning on the original data set, deleting invalid data points and interpolating missing data to generate a cleaned multi-source meteorological data set;

[0012] S3. Manually label the cleaned multi-source meteorological dataset to generate a multi-source meteorological training sample set, wherein the training sample set includes type information of weather systems and corresponding spatial location labels;

[0013] S4. Based on the training sample set, multiple single-source meteorological feature extraction models are trained respectively, each model is built based on the backbone part of the YOLOv5 network structure, takes the cleaned single meteorological data as input, and the weather system label as output, to obtain corresponding single-source feature output;

[0014] S5. Concatenate the single-source feature outputs according to the channel dimension to obtain a fused feature tensor;

[0015] S6. Inputting the fused feature tensor into an integrated recognition neural network composed of a Neck and Head structure for training to obtain a multi-source integrated weather system recognition model;

[0016] S7. The real-time multi-source meteorological data of the area to be identified is processed according to step S2 and sequentially input into the trained single-source meteorological feature extraction model and the multi-source integrated recognition model, and an identification result including the weather system type and location coordinates is output.

[0017] Optionally, the multi-source meteorological data in step S1 include ground wind speed field, relative humidity field, sea level pressure field, visible light and infrared channel cloud map images obtained by geostationary orbit satellites, and radar reflectivity puzzle images generated by ground-based Doppler weather radar. All multi-source meteorological data are preprocessed and rasterized according to a unified time resolution and spatial resolution to construct an original meteorological data set with a consistent structure that can be jointly input into the model.

[0018] Optionally, the data cleaning process in step S2 includes the following operations: removing abnormal values ​​in the original meteorological data, the abnormal values ​​include observation values ​​with values ​​of 9999, NaN or other non-physical meanings, and using interpolation methods to supplement missing raster data, the interpolation methods include linear interpolation and nearest neighbor interpolation, and the data obtained after processing maintains consistency with the original data in terms of time axis and spatial resolution.

[0019] Optionally, step S3 specifically includes: slicing the preprocessed multi-source meteorological dataset at the frame level in chronological order, dividing it into data frames at corresponding time points, manually interpreting each frame of the image, and combining the wind speed field, humidity field, pressure field, satellite cloud map and radar reflectivity map to mark the type information of the existing weather system and its corresponding spatial position, the type information includes convective system, cold front system and tropical cyclone system, and the spatial position includes the center position and boundary range of the system in the image, and forming a training sample pair of each frame of the image and its corresponding annotation result to construct a multi-source meteorological training sample set.

[0020] Optionally, the S4 specifically includes:

[0021] S41, the training sample set D constructed for multi-source meteorological data is {(F1, L1), (F2, L2), ..., (F T ,L T)}, split the multi-source meteorological data according to the meteorological element type, and extract each type of meteorological element data into a single dataset, forming a wind speed dataset, a humidity dataset, an air pressure dataset, a satellite image dataset, and a radar image dataset;

[0022] S42. Construct a corresponding feature extraction model for each type of single meteorological data. The model structure is the Backbone part of the YOLOv5 neural network, and the input tensor size is received. in, represents the input image of the kth type of meteorological data in the i-th frame, H and W are the image height and width respectively, C k is the number of channels;

[0023] S43, each type of single meteorological data frame With the corresponding label L i Together they are input into the kth feature extraction model for supervised training, so that the output feature tensor can effectively encode the weather system type c i,j and spatial positioning (x i,j ,y i,j ,w i,j ,h i,j ) related discriminative features;

[0024] S44, through multiple rounds of iterative training, K single-source meteorological feature extraction models {M1, M2, ..., M K}, where K = 5, corresponding to five types of data: wind speed, humidity, air pressure, satellite imagery, and radar imagery. Each model outputs its corresponding feature representation. That is, the Backbone output of the k-th type of data in the i-th frame;

[0025] S45. Concatenate the output feature tensors of all single-source models according to the channel dimension to generate the feature tensor set required for fusion input

[0026] Optionally, the S5 specifically includes:

[0027] S51. Set of fused feature tensors The integrated input tensor is constructed by channel splicing. The channel dimension is the sum of the output channels of each single-source feature extraction model to form an input set.

[0028] S52, construct an integrated recognition neural network based on the Neck and Head parts of the YOLOv5 network structure, the network is based on As input, the frame label set L i To supervise the output, the network structure includes a feature fusion module, a path enhancement module and a target detection output module;

[0029] S53, input the fusion features in the training sample set Each of them is input into the integrated recognition neural network, and forward propagation and reverse gradient update are performed to optimize the loss function, which includes the sum of multi-class cross entropy loss and the IoU loss between the predicted bounding box and the true bounding box;

[0030] S54, through batch training and multiple rounds of iteration, in the training sample set The network parameters convergence is completed, and finally a complete multi-source integrated weather system identification model is formed, which has the ability to predict the type and location of weather systems based on the input multi-source meteorological data.

[0031] Optionally, the S6 specifically includes:

[0032] S61, obtaining multi-source meteorological data of the area to be identified at the current moment, the data including wind speed, humidity, sea level pressure, satellite images, and radar images;

[0033] S62: Preprocess the current multi-source meteorological data, including noise removal and missing completion, to generate a cleaned meteorological data frame.

[0034] S63, each type of cleaned single source meteorological data frame Input the corresponding single-source meteorological feature extraction model M respectively k , extract Backbone feature tensor A total of five single-source feature tensors are obtained;

[0035] S64, for five types of feature tensors Perform channel dimension splicing to generate fusion feature tensor

[0036] S65, fusion feature tensor Input the trained multi-source integrated weather system recognition model, perform forward reasoning, and output a set of recognition results:

[0037]

[0038] in, represents the center position of the jth weather system, Indicates the border size, Indicates the type of weather system identified, is the number of weather systems identified in the current frame.

[0039] Optionally, each item in the identification result set in step S65 In addition to the spatial location information and type number of the weather system, it also contains the system confidence score, which indicates the reliability of the model's judgment on the recognition result. The value range is 0 to 1. The final output results are sorted from high to low according to the confidence.

[0040] Optionally, the S7 specifically includes: loading the trained multi-source integrated weather system recognition model into the reasoning environment, the reasoning environment supports the image input interface and the recognition result structured output interface, initializing the corresponding model parameters and reasoning logic, receiving the area to be recognized at any given time point The corresponding pre-processed multi-source meteorological data Each type of meteorological data is automatically input into the corresponding single-source feature extraction model M through the built-in data scheduling module. k , extract the feature tensor Perform channel concatenation on all feature tensors to form a fused feature tensor Input into the loaded integrated recognition model to perform a complete forward reasoning, and output a set of recognition results including category, location and confidence For the recognition result set Confidence screening is performed on each item in the dataset, recognition results with confidence levels below the set threshold are eliminated, and the retained results are converted into structured output formats, including JSON, GeoJSON, or Shapefile.

[0041] The beneficial effects of the present invention are:

[0042] (1) This invention introduces multi-source heterogeneous meteorological data such as wind speed, humidity, sea level pressure, satellite images, and radar images into the same recognition process for the first time. By constructing a single-source feature extraction model, each type of meteorological element is modeled separately and then fused for recognition, thus avoiding the feature interference problem caused by mixed input in traditional methods. While maintaining structural clarity, the model improves the system's ability to discriminate weather systems and perceive spatiotemporal details.

[0043] (2) Based on the YOLOv5 neural network structure, this paper restructures the Backbone, Neck, and Head modules and introduces a training mechanism of "single-source extraction first, centralized fusion later," achieving complete link optimization from multi-source feature acquisition to integrated reasoning. This layered architecture not only effectively improves the model's adaptability to weather systems of different scales, but also facilitates the introduction of new meteorological data or replacement of some model modules in the future, and has good system scalability and engineering versatility.

[0044] (3) During the model application phase, this paper designs a structured reasoning output and confidence screening mechanism. The output results include location coordinates, system type, and confidence score, and can be directly generated into standard formats such as JSON or GeoJSON, facilitating integration with meteorological warning platforms and geographic information visualization systems. Furthermore, this method possesses the advantages of real-time processing capabilities and lightweight model deployment, making it suitable for high-frequency, automated operation in meteorological stations, local servers, or cloud-edge collaborative platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0046] Figure 1 This is a flow chart of a weather system identification method based on multi-source meteorological information proposed by the present invention;

[0047] Figure 2 This is a model structure diagram of a weather system identification method based on multi-source meteorological information proposed in the present invention. DETAILED DESCRIPTION

[0048] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0049] refer to Figure 1 and Figure 2 A weather system identification method based on multi-source meteorological information includes the following steps:

[0050] S1. Obtain meteorological data from multiple sources to form an original data set;

[0051] S2. Performing data cleaning on the original data set, deleting invalid data points and interpolating missing data to generate a cleaned multi-source meteorological data set;

[0052] S3. Manually label the cleaned multi-source meteorological dataset to generate a multi-source meteorological training sample set, wherein the training sample set includes type information of weather systems and corresponding spatial location labels;

[0053] S4. Based on the training sample set, multiple single-source meteorological feature extraction models are trained respectively, each model is built based on the backbone part of the YOLOv5 network structure, takes the cleaned single meteorological data as input, and the weather system label as output, to obtain corresponding single-source feature output;

[0054] S5. Concatenate the single-source feature outputs according to the channel dimension to obtain a fused feature tensor;

[0055] S6. Inputting the fused feature tensor into an integrated recognition neural network composed of a Neck and Head structure for training to obtain a multi-source integrated weather system recognition model;

[0056] S7. The real-time multi-source meteorological data of the area to be identified is processed according to step S2 and sequentially input into the trained single-source meteorological feature extraction model and the multi-source integrated recognition model, and an identification result including the weather system type and location coordinates is output.

[0057] This invention constructs a weather system identification method based on multi-source meteorological information, effectively integrating five core meteorological elements: wind speed, humidity, air pressure, satellite imagery, and radar charts. This method establishes a complete identification process encompassing data collection, cleaning, feature extraction, integrated training, and inference output. Compared to traditional methods that rely on a single data source or purely manual interpretation, this method boasts stronger information integration capabilities, a higher degree of automation, and superior identification stability, making it widely applicable to high-precision weather warning systems and intelligent weather service systems.

[0058] In this embodiment, the multi-source meteorological data in step S1 include ground wind speed field, relative humidity field, sea level pressure field, visible light and infrared channel cloud map images obtained by geostationary satellites, and radar reflectivity puzzle images generated by ground-based Doppler weather radar. All multi-source meteorological data are preprocessed and rasterized according to a unified time resolution and spatial resolution to construct an original meteorological data set with a consistent structure that can be jointly input into the model.

[0059] The present invention clearly defines the composition and processing structure of multi-source meteorological data, including the collaborative organization of numerical fields such as wind speed, humidity, and sea level pressure and remote sensing image data. It solves the problem of inconsistency between different data formats, resolutions, and time steps in the existing technology, ensures the uniformity of subsequent network input structure and training stability, and significantly improves data utilization efficiency and model input quality.

[0060] In this embodiment, the data cleaning process in step S2 includes the following operations: removing abnormal values ​​in the original meteorological data, the abnormal values ​​include observation values ​​with a value of 9999, NaN or other non-physical meanings, and using interpolation methods to supplement missing raster data. The interpolation methods include linear interpolation and nearest neighbor interpolation. The data obtained after processing maintains consistency with the original data in terms of time axis and spatial resolution.

[0061] This paper systematically defines the data preprocessing steps and proposes a solution for removing and completing typical outliers and missing items in meteorological data. By using linear interpolation and nearest neighbor interpolation to improve data continuity, it effectively addresses information interruptions caused by sensor omissions and observation anomalies. This provides structurally complete and stable input for model training and inference, enhancing overall system robustness.

[0062] In this embodiment, step S3 specifically includes: slicing the preprocessed multi-source meteorological data set at the frame level in chronological order, dividing it into data frames at corresponding time points, manually interpreting each frame image, and combining the wind speed field, humidity field, pressure field, satellite cloud map and radar reflectivity map to mark the type information of the existing weather system and its corresponding spatial position. The type information includes convective system, cold front system and tropical cyclone system, and the spatial position includes the center position and boundary range of the system in the image. Each frame image and its corresponding annotation result are combined into a training sample pair to construct a multi-source meteorological training sample set.

[0063] Through high-quality manual labeling by professional meteorologists, this invention establishes a refined training sample set, clarifies the types of weather systems and the corresponding spatial boundaries, and is significantly superior to traditional threshold discrimination or automatic reasoning to generate labels. The labeled information has higher accuracy and interpretability, providing a solid training foundation for deep learning models and ensuring that the recognition results are close to actual meteorological cognition.

[0064] In this embodiment, the S4 specifically includes:

[0065] S41, the training sample set D constructed for multi-source meteorological data is {(F1, L1), (F2, L2), ..., (F T ,L T )}, split the multi-source meteorological data according to the meteorological element type, and extract each type of meteorological element data into a single dataset, forming a wind speed dataset, a humidity dataset, an air pressure dataset, a satellite image dataset, and a radar image dataset;

[0066] S42. Construct a corresponding feature extraction model for each type of single meteorological data. The model structure is the Backbone part of the YOLOv5 neural network, and the input tensor size is received. in, represents the input image of the kth type of meteorological data in the i-th frame, H and W are the image height and width respectively, C k is the number of channels;

[0067] S43, each type of single meteorological data frame With the corresponding label L iTogether they are input into the kth feature extraction model for supervised training, so that the output feature tensor can effectively encode the weather system type c i,j and spatial positioning (x i,j ,y i,j ,w i,j ,h i,j ) related discriminative features;

[0068] S44, through multiple rounds of iterative training, K single-source meteorological feature extraction models {M1, M2, ..., M K}, where K = 5, corresponding to five types of data: wind speed, humidity, air pressure, satellite imagery, and radar imagery. Each model outputs its corresponding feature representation. That is, the Backbone output of the k-th type of data in the i-th frame;

[0069] S45. Concatenate the output feature tensors of all single-source models according to the channel dimension to generate the feature tensor set required for fusion input

[0070] This paper establishes an independent training process for a single-source feature extraction model and proposes a "source-specific modeling and structural decoupling" approach. Compared to the traditional multi-channel direct splicing input method, this method significantly improves the ability to express independent features of various meteorological elements, avoids feature conflicts and information interference, provides a more stable and controllable feature input space for multi-source fusion, and enhances the model's generalization and structural flexibility.

[0071] In this embodiment, the S5 specifically includes:

[0072] S51. Set of fused feature tensors The integrated input tensor is constructed by channel splicing. The channel dimension is the sum of the output channels of each single-source feature extraction model to form an input set.

[0073] S52, construct an integrated recognition neural network based on the Neck and Head parts of the YOLOv5 network structure, the network is based on As input, the frame label set L i To supervise the output, the network structure includes a feature fusion module, a path enhancement module and a target detection output module;

[0074] S53, input the fusion features in the training sample set Each of them is input into the integrated recognition neural network, and forward propagation and reverse gradient update are performed to optimize the loss function, which includes the sum of multi-class cross entropy loss and the IoU loss between the predicted bounding box and the true bounding box;

[0075] S54, through batch training and multiple rounds of iteration, in the training sample set The network parameters convergence is completed, and finally a complete multi-source integrated weather system identification model is formed, which has the ability to predict the type and location of weather systems based on the input multi-source meteorological data.

[0076] This paper reconstructs the Neck and Head structures of YOLOv5 and designs an integrated recognition network architecture for multi-source feature fusion. By introducing a structured training path and a combined loss function, the model's classification ability and positioning accuracy for different types of weather systems are enhanced. This enables end-to-end recognition modeling after deep fusion, improving the overall model's detection accuracy and recognition efficiency.

[0077] In this embodiment, S6 specifically includes:

[0078] S61, obtaining multi-source meteorological data of the area to be identified at the current moment, the data including wind speed, humidity, sea level pressure, satellite images, and radar images;

[0079] S62: Preprocess the current multi-source meteorological data, including noise removal and missing completion, to generate a cleaned meteorological data frame.

[0080] S63, each type of cleaned single source meteorological data frame Input the corresponding single-source meteorological feature extraction model M respectively k , extract Backbone feature tensor A total of five single-source feature tensors are obtained;

[0081] S64, for five types of feature tensors Perform channel dimension splicing to generate fusion feature tensor

[0082] S65, fusion feature tensor Input the trained multi-source integrated weather system recognition model, perform forward reasoning, and output a set of recognition results:

[0083]

[0084] in, represents the center position of the jth weather system, Indicates the border size, Indicates the type of weather system identified, is the number of weather systems identified in the current frame.

[0085] This invention enables the deployment of trained models and the inference application of real-time meteorological data, proposing a closed-loop mechanism from data acquisition, preprocessing, feature extraction, to inference output. This mechanism ensures data structure consistency between the training and inference phases, effectively supporting the operational stability of the model in real-world meteorological environments, and exhibiting excellent timeliness, scalability, and interface compatibility.

[0086] In this embodiment, each item in the identification result set in step S65 In addition to the spatial location information and type number of the weather system, it also contains the system confidence score, which indicates the reliability of the model's judgment on the recognition result. The value range is 0 to 1. The final output results are sorted from high to low according to the confidence.

[0087] This paper introduces a confidence screening mechanism that sets confidence thresholds for recognition output results, improving the model's controllability and ensuring the quality of results at the output layer. This mechanism helps filter out low-confidence recognition targets and highlight high-reliability warning areas, enhancing the system's practical decision-making support value in meteorological business scenarios and supporting hierarchical display of results and post-processing analysis.

[0088] In this embodiment, the S7 specifically includes: loading the trained multi-source integrated weather system recognition model into the reasoning environment, the reasoning environment supports the image input interface and the recognition result structured output interface, initializing the corresponding model parameters and reasoning logic, receiving the area to be recognized at any given time point, The corresponding pre-processed multi-source meteorological data Each type of meteorological data is automatically input into the corresponding single-source feature extraction model M through the built-in data scheduling module. k , extract the feature tensor Perform channel concatenation on all feature tensors to form a fused feature tensor Input into the loaded integrated recognition model to perform a complete forward reasoning, and output a set of recognition results including category, location and confidence For the recognition result set Confidence screening is performed on each item in the dataset, recognition results with confidence levels below the set threshold are eliminated, and the retained results are converted into structured output formats, including JSON, GeoJSON, or Shapefile.

[0089] This invention specifies the inference system's operating environment, model loading process, interface input and output specifications, and result formatting mechanism in detail, forming a complete business-level deployment solution. Compared to traditional research-based recognition models, this invention boasts stronger engineering capabilities and system integration. Recognition results can be directly used in visualization platforms or accessed through early warning systems, significantly improving the model's usability and practical value.

[0090] Example 1:

[0091] To verify the feasibility and recognition effectiveness of this invention in actual business applications, we selected the jurisdiction of a prefecture-level city meteorological bureau in East my country as the test area. The test time was 8:00 to 9:00 a.m. on July 1, 2023, and the test range included the period of high incidence of severe convection. In the early morning of that day, the area showed significant precursory characteristics such as increased wind speed, increased humidity, and gradually strengthened radar echoes, which are typical development scenarios for short-term heavy rainfall. Traditional manual analysis methods require multiple operators to cross-check between radar stations and business platforms, while numerical models cannot provide real-time recognition results due to computational delays.

[0092] In this scenario, we deployed the multi-source meteorological information recognition model system proposed in this paper. The system's front-end acquisition module accesses ERA5 wind speed, humidity, and air pressure products, and simultaneously samples them in conjunction with local Doppler radar and geostationary satellite data. The collected data is fed into five single-source feature extraction models (based on the YOLOv5 Backbone) in real time. A fusion module then combines these data to form a unified feature tensor, which is then fed into an integrated recognition model for inference. With an inference latency of less than 5 seconds, the model accurately identifies the development trend of convective systems and outputs the recognition category, boundary information, and confidence score in real time.

[0093] The identification results show that between 8:10 and 8:40, the model continuously output convective system signals with a high confidence level. As the echo intensity exceeded 40 dBZ, the system automatically identified the area as a heavy rainfall risk area and generated structured GeoJSON for display on downstream platforms. Compared to manual post-event analysis, the system's identification was 18 minutes ahead of schedule, with a spatial positioning error of less than 5 kilometers, fully meeting the meteorological department's operational requirements for "early warning and rapid response."

[0094] Table 1: Thunderstorm system identification and verification experimental data

[0095]

[0096] The data in the table shows that meteorological conditions in the target area underwent significant changes between 08:00 and 08:40 on July 1, 2023. Wind speeds gradually increased from 5.4 to 8.2 m / s, reflecting an active surface wind field. Humidity continued to rise from 82% to 91%, suggesting that atmospheric water vapor conditions were gradually approaching saturation. Simultaneously, air pressure steadily decreased from 1003.2 hPa to 1001.0 hPa, indicating increased control by the low-pressure system. More critically, the radar reflectivity (dBZ) rose rapidly from 23.5 to 41.7, entering the critical range for severe convection and indicating significant convection development with potential for disasters.

[0097] During this process, the multi-source meteorological identification system constructed by the present invention responded quickly. At 08:00, the system identification result was "no system", and the confidence level was only 0.31, indicating that the model did not make any false alarms. As the radar echo increased to 28.1dBZ at 08:10, the model recognized the "appearance of a convective system" for the first time, and the confidence level increased to 0.65. Subsequently, at 08:20 and 08:30, the system continuously judged that "the convective system is strengthening" and "the system is significantly strengthened", and the recognition confidence levels rose to 0.78 and 0.88 respectively, indicating that the model has a good tracking capability for the system evolution process. By 08:40, the radar echo reached 41.7dBZ, and the system accurately gave the judgment of "identifying the risk of heavy rainfall" with a confidence level of up to 0.93.

[0098] The entire analysis process demonstrates that this invention not only effectively leverages multi-source meteorological information to generate early-stage response judgments for weather emergencies, but also dynamically adjusts recognition results based on changing trends in meteorological factors. Its confidence scoring mechanism effectively reflects the evolution of recognition signal strength. By analyzing the linkage between time, meteorological factors, and system output, this table clearly demonstrates the comprehensive advantages of this invention in real-time recognition accuracy, response sensitivity, and false alarm suppression, demonstrating its excellent practicality and engineering value.

[0099] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A weather system identification method based on multi-source meteorological information, characterized in that: The steps include: S1. Obtain meteorological data from multiple sources to form an original data set; S2. Performing data cleaning on the original data set, deleting invalid data points and interpolating missing data to generate a cleaned multi-source meteorological data set; S3. Manually label the cleaned multi-source meteorological dataset to generate a multi-source meteorological training sample set, wherein the training sample set includes type information of weather systems and corresponding spatial location labels; S4. Based on the training sample set, multiple single-source meteorological feature extraction models are trained respectively, each model is built based on the backbone part of the YOLOv5 network structure, takes the cleaned single meteorological data as input, and the weather system label as output, to obtain corresponding single-source feature output; S5. Concatenate the single-source feature outputs according to the channel dimension to obtain a fused feature tensor; S6. Inputting the fused feature tensor into an integrated recognition neural network composed of a Neck and Head structure for training to obtain a multi-source integrated weather system recognition model; S7. The real-time multi-source meteorological data of the area to be identified is processed according to step S2 and sequentially input into the trained single-source meteorological feature extraction model and the multi-source integrated recognition model, and an identification result including the weather system type and location coordinates is output.

2. A weather system identification method based on multi-source meteorological information according to claim 1, characterized in that: The multi-source meteorological data in step S1 include ground wind speed field, relative humidity field, sea level pressure field, visible light and infrared channel cloud map images obtained by geostationary orbit satellites, and radar reflectivity mosaic images generated by ground-based Doppler weather radar. All multi-source meteorological data are preprocessed and rasterized according to a unified temporal resolution and spatial resolution to construct an original meteorological data set with a consistent structure that can be jointly input into the model.

3. The weather system identification method based on multi-source meteorological information according to claim 1, characterized in that: The data cleaning process in step S2 includes the following operations: removing abnormal values ​​in the original meteorological data, wherein the abnormal values ​​include observation values ​​with a value of 9999, NaN, or other non-physical meanings, and using interpolation methods to supplement missing raster data, wherein the interpolation methods include linear interpolation and nearest neighbor interpolation. The data obtained after processing maintains consistency with the original data in terms of time axis and spatial resolution.

4. The weather system identification method based on multi-source meteorological information according to claim 1, characterized in that: The step S3 specifically includes: slicing the preprocessed multi-source meteorological dataset at the frame level in chronological order, dividing it into data frames at corresponding time points, manually interpreting each frame of the image, and combining the wind speed field, humidity field, pressure field, satellite cloud image and radar reflectivity image to mark the type information of the existing weather system and its corresponding spatial position, wherein the type information includes convective system, cold front system and tropical cyclone system, and the spatial position includes the center position and boundary range of the system in the image, and each frame of the image and its corresponding annotation result are combined into a training sample pair to construct a multi-source meteorological training sample set.

5. The weather system identification method based on multi-source meteorological information according to claim 1, characterized in that: The S4 specifically includes: S41, the training sample set D constructed for multi-source meteorological data is {(F1, L1), (F2, L2), ..., (F T ,L T )}, split the multi-source meteorological data according to the meteorological element type, and extract each type of meteorological element data into a single dataset, forming a wind speed dataset, a humidity dataset, an air pressure dataset, a satellite image dataset, and a radar image dataset; S42. Construct a corresponding feature extraction model for each type of single meteorological data. The model structure is the Backbone part of the YOLOv5 neural network, and the input tensor size is received. in, represents the input image of the kth type of meteorological data in the i-th frame, H and W are the image height and width respectively, C k is the number of channels; S43, each type of single meteorological data frame With the corresponding label L i Together they are input into the kth feature extraction model for supervised training, so that the output feature tensor can effectively encode the weather system type c i,j and spatial positioning (x i,j ,y i,j ,w i,j ,h i,j ) related discriminative features; S44, through multiple rounds of iterative training, K single-source meteorological feature extraction models {M1, M2, ..., M K }, where K = 5, corresponding to five types of data: wind speed, humidity, air pressure, satellite imagery, and radar imagery. Each model outputs its corresponding feature representation. That is, the Backbone output of the k-th type of data in the i-th frame; S45. Concatenate the output feature tensors of all single-source models according to the channel dimension to generate the feature tensor set required for fusion input 6. The weather system identification method based on multi-source meteorological information according to claim 1, characterized in that: The S5 specifically includes: S51. Set of fused feature tensors The integrated input tensor is constructed by channel splicing. The channel dimension is the sum of the output channels of each single-source feature extraction model to form an input set. S52, construct an integrated recognition neural network based on the Neck and Head parts of the YOLOv5 network structure, the network is based on As input, the frame label set L i To supervise the output, the network structure includes a feature fusion module, a path enhancement module and a target detection output module; S53, input the fusion features in the training sample set Each of them is input into the integrated recognition neural network, and forward propagation and reverse gradient update are performed to optimize the loss function, which includes the sum of multi-class cross entropy loss and the IoU loss between the predicted bounding box and the true bounding box; S54, through batch training and multiple rounds of iteration, in the training sample set The network parameters convergence is completed, and finally a complete multi-source integrated weather system identification model is formed, which has the ability to predict the type and location of weather systems based on the input multi-source meteorological data.

7. The weather system identification method based on multi-source meteorological information according to claim 1, characterized in that: The S6 specifically includes: S61, obtaining multi-source meteorological data of the area to be identified at the current moment, the data including wind speed, humidity, sea level pressure, satellite images, and radar images; S62: Preprocess the current multi-source meteorological data, including noise removal and missing completion, to generate a cleaned meteorological data frame. S63, each type of cleaned single source meteorological data frame Input the corresponding single-source meteorological feature extraction model M respectively k , extract Backbone feature tensor A total of five single-source feature tensors are obtained; S64, for five types of feature tensors Perform channel dimension splicing to generate fusion feature tensor S65, fusion feature tensor Input the trained multi-source integrated weather system recognition model, perform forward reasoning, and output a set of recognition results: in, represents the center position of the jth weather system, Indicates the border size, Indicates the type of weather system identified, is the number of weather systems identified in the current frame.

8. The weather system identification method based on multi-source meteorological information according to claim 7, characterized in that: In step S65, each item in the identification result set is identified In addition to the spatial location information and type number of the weather system, it also contains the system confidence score, which indicates the reliability of the model's judgment on the recognition result. The value range is 0 to 1. The final output results are sorted from high to low according to the confidence.

9. The weather system identification method based on multi-source meteorological information according to claim 1, characterized in that: The S7 specifically includes: loading the trained multi-source integrated weather system recognition model into the reasoning environment, the reasoning environment supports an image input interface and a structured output interface for recognition results, initializing the corresponding model parameters and reasoning logic, and receiving the image data of the area to be recognized at any given time point. The corresponding pre-processed multi-source meteorological data Each type of meteorological data is automatically input into the corresponding single-source feature extraction model M through the built-in data scheduling module. k , extract the feature tensor Perform channel concatenation on all feature tensors to form a fused feature tensor Input into the loaded integrated recognition model to perform a complete forward reasoning, and output a set of recognition results including category, location and confidence For the recognition result set Confidence screening is performed on each item in the dataset, recognition results with confidence levels below the set threshold are eliminated, and the retained results are converted into structured output formats, including JSON, GeoJSON, or Shapefile.

Images