Severe convection monitoring method and system based on multi-source data

Abstract

The invention relates to a severe convection monitoring method and system based on multi-source data. The method comprises the following steps: collecting multi-source meteorological data; preprocessing the multi-source meteorological data to obtain a preprocessed meteorological data set; on each space grid unit, extracting a multi-source characteristic parameter used for reflecting the severe convection activity from the preprocessed meteorological data set; inputting the multi-source characteristic parameters into a pre-trained scoring model for reasoning analysis, and generating a severe convection score value of each space grid unit; performing fuzzy classification based on the severe convection score value of each space grid unit and a preset fuzzy membership rule to obtain a severe convection grade result of each space grid unit; and calculating the convection development trend of each space grid unit based on the numerical difference between the severe convection score value and the continuous score value at the previous two moments. The method has the effect of improving the precision of severe convection monitoring.

Description

Technical Field

[0001] This application relates to the technical field of meteorological monitoring and disaster early warning, and in particular to a method and system for monitoring severe convection based on multi-source data. Background Technology

[0002] Currently, severe convective weather, as a high-impact short-term meteorological disaster, is characterized by its suddenness and localization, often accompanied by thunderstorms, strong winds, hail, and short-duration heavy rainfall, seriously threatening people's lives and property. Real-time monitoring and development trend assessment of severe convective activity are crucial foundations for achieving short-term forecasting and early warning.

[0003] Existing severe convection monitoring technologies mainly rely on single data sources, such as ground-based weather radar, geostationary satellite imagery, or numerical weather prediction products. These methods are often based on fixed thresholds or rule systems to determine whether a certain type of observational indicator (such as radar reflectivity exceeding a certain value) meets the conditions for severe convection. However, single-source indicators are easily affected by factors such as local topography, sensor blind spots, or missing data, and cannot comprehensively reflect the combined processes of atmospheric instability, uplift triggering, and convection development, making them prone to misjudgment or omission.

[0004] The existing technical solutions mentioned above have the following drawbacks: most existing severe convective weather monitoring methods adopt static rules, single-source input and fixed classification methods, and lack deep fusion of multi-source heterogeneous data, which leads to inaccurate level determination in complex meteorological scenarios, and therefore there is room for improvement. Summary of the Invention

[0005] To improve the accuracy of severe convection monitoring, this application provides a method and system for severe convection monitoring based on multi-source data.

[0006] The above-mentioned objective of this application is achieved through the following technical solution: A method for monitoring severe convection based on multi-source data, characterized in that the method includes: Collect multi-source meteorological data, including radar data, satellite image data, lightning detection data, ground meteorological station data, radiosonde profile data, and numerical weather model products; The multi-source meteorological data is preprocessed to obtain a preprocessed meteorological dataset, and the preprocessed meteorological dataset is mapped to multiple spatial grid units according to a preset spatial division rule. Each spatial grid unit corresponds to a fixed geographical area and observation time interval. In each spatial grid cell, multi-source feature parameters reflecting severe convective activity are extracted from the preprocessed meteorological dataset. These multi-source feature parameters include radar reflectivity, cloud top brightness temperature, lightning frequency, CAPE, and wind shear. The multi-source feature parameters are input into a pre-trained scoring model for inference and analysis to generate a strong convection score value for each spatial grid cell. The strong convection score value is a continuous real number in the range of 0 to 1, which is used to reflect the risk level of strong convection activity in the corresponding cell and is used for subsequent fuzzy classification level judgment. Fuzzy classification is performed based on the strong convection score of each spatial grid cell and the preset fuzzy membership rules to obtain the strong convection level result of each spatial grid cell. The strong convection level classification includes no convection, weak convection, moderate convection and strong convection. Based on the numerical difference between the strong convection score and the consecutive scores of the previous two moments, the convection development trend of each spatial grid cell is calculated, and the convection development trend includes three states: enhancement, weakening or stabilization. By combining the strong convection level results and convection development trend status of each spatial grid cell, a multi-resolution convection monitoring layer is constructed, which includes a risk level map, a development trend map, and a confidence level map.

[0007] By adopting the above technical solutions, and collecting multi-source meteorological data, it is possible to comprehensively acquire various types of meteorological information reflecting convective activity, thereby enhancing the collaborative perception capability of different meteorological elements. Preprocessing the multi-source meteorological data and mapping it to multiple spatial grid cells unifies the spatiotemporal scale of the data, ensuring the accuracy and consistency of subsequent feature extraction and analysis processes, thus improving the adaptability and operational efficiency of the analysis model. Extracting multi-source feature parameters from each spatial grid cell and inputting them into a scoring model for inference analysis dynamically generates physically meaningful severe convective weather scores, enabling a quantitative assessment of the risk of severe convective weather occurrence. By performing fuzzy classification and trend judgment on the scores, the convective level and its evolution direction can be accurately identified, thereby improving the lead time and reliability of disaster warnings. Constructing multi-resolution convective monitoring layers supports multi-level display and interactive analysis, thereby enhancing the system's practical value in forecasting, early warning, and emergency response scenarios.

[0008] In one example, this application can be further configured such that the collection of multi-source meteorological data specifically includes: Acquire radar observation data covering the monitoring area, the radar observation data including three-dimensional reflectivity factor, radial velocity and echo top height; Acquire multispectral remote sensing image data from geostationary satellites, including infrared channel brightness temperature, visible light albedo, cloud top brightness temperature, and cloud type identification results. Acquire real-time lightning detection data from the lightning location system, including ground lightning type, occurrence time, geographical location, and frequency data; Collect observation data from ground-based automatic weather stations, including air temperature, relative humidity, surface wind speed, air pressure, and hourly precipitation. Analyze the vertical profile data provided by the upper-air sounding system, which includes temperature profile, humidity profile, wind speed and direction, vertical structure and stability parameters; The analysis field and forecast products of the numerical weather model are invoked, including convective available potential energy, wind shear, vertical velocity, isentropic potential temperature, and convective parameter fields.

[0009] By adopting the above technical solutions and acquiring multi-source meteorological data, including radar, satellite, lightning, ground observation, radiosonde, and numerical models, it is possible to comprehensively cover the key physical processes of the formation, development, and triggering conditions of convective systems, thereby constructing a complete information perception chain and improving the accuracy and robustness of severe convective monitoring.

[0010] In one example, this application can be further configured such that: the preprocessing of the multi-source meteorological data to obtain a preprocessed meteorological dataset specifically includes: Spatial interpolation and grid resampling operations are performed on radar data, satellite image data and numerical model products in the multi-source meteorological data to uniformly map all types of data to a preset spatial grid structure. Time alignment processing is performed on various types of observation data. The time alignment processing includes performing time resampling and time interpolation operations on data with different observation time intervals to unify them into a fixed observation time interval. Time-series interpolation is used to complete missing data in lightning detection data and ground meteorological station data in the multi-source meteorological data, and hierarchical reconstruction is performed on radiosonde profile data in the multi-source meteorological data to ensure profile integrity. The processed data is formatted and integrated to generate the preprocessed meteorological dataset, which is then mapped to the corresponding spatial grid cells. The preprocessed meteorological dataset includes spatial coordinate indexes, observation time labels, and multi-source feature fields.

[0011] By adopting the above technical solutions, and by performing spatial interpolation, time alignment, missing data completion, and formatting on different data sources, the original multi-source heterogeneous meteorological data can be transformed into a standardized meteorological dataset with a unified structure. This ensures the standardization of inputs required for subsequent model inference and improves the stability and engineering feasibility of the overall processing flow.

[0012] In one example, this application can be further configured such that: the extraction of multi-source feature parameters reflecting severe convective activity from the preprocessed meteorological dataset specifically includes: Based on the preset convection physics mechanism and feature extraction rules, the maximum reflectivity, echo top height and vertical gradient of radial velocity at the current moment are extracted from the radar observation data in the preprocessed meteorological dataset in each spatial grid cell. Cloud top brightness temperature, brightness temperature change rate, and cloud type category index based on brightness temperature determination are extracted from the satellite image data in the preprocessed meteorological dataset. The frequency of ground flashes and the historical cumulative lightning density per unit time are calculated from the lightning detection data in the preprocessed meteorological dataset. Near-surface air temperature, humidity and wind speed are extracted from the ground weather station data in the preprocessed meteorological dataset, and the ground disturbance intensity is calculated by combining the difference between adjacent grids. The convective potential energy, convective suppression energy, wind shear index, and vertical humidity gradient are extracted from the radiosonde profile data in the preprocessed meteorological dataset. The vertical velocity, isentropic potential temperature field, and 500 hPa layer potential temperature anomaly values ​​within the forecast period are obtained from the numerical weather model products in the preprocessed meteorological dataset to assist in determining mid-to-high-level instability and uplift triggering conditions.

[0013] By adopting the above technical solutions, and by extracting convection-related features from various types of data based on convection physics mechanisms and preset rules, it is possible to accurately mine key indicators reflecting instability, uplift triggering, and electromagnetic activity, thereby improving the pertinence of strong convection identification and the physical interpretability of model predictions.

[0014] In one example, this application can be further configured such that: the method for monitoring strong convection based on multi-source data also includes: A training sample set is constructed based on historical severe convection samples and corresponding spatiotemporally matched multi-source meteorological feature data. The training samples include multi-source feature vectors of spatial grid cells and severe convection occurrence labels labeled with real-time data. The training samples are trained using a multilayer perceptron neural network. The model input is a feature vector composed of multi-source feature parameters, and the model output is a continuous strong convection risk score value between 0 and 1. During model training, the cross-entropy loss function is used as the optimization objective, and the model overfitting is controlled by early stopping strategy and validation set evaluation. The trained model is used as the pre-trained scoring model.

[0015] By adopting the above technical solution and constructing a spatiotemporal paired training set of strong convection real-time labels and multi-source meteorological data, risk patterns in historical samples can be effectively captured, thus providing a reliable learning basis for the scoring model. By training the neural network scoring model and optimizing its generalization ability, stable and discriminative risk scoring results can be output, thus providing a reliable basis for classification and trend analysis.

[0016] In one example, this application can be further configured as follows: inputting the multi-source feature parameters into a pre-trained scoring model for inference analysis to generate a strong convection score value for each spatial grid cell, specifically including: In each spatial grid cell, a standardized feature vector is constructed based on the multi-source feature parameters; The standardized feature vector is input into the pre-trained scoring model, and forward inference is performed to output a strong convection score value between 0 and 1, which is used to characterize the risk level of strong convection activity in the current grid cell. During the reasoning process, a dynamic weight adjustment mechanism is introduced based on the data integrity and scene adaptability of each data source within the current grid cell. This mechanism dynamically adjusts the weights of each dimension of the standardized feature vector to improve the stability and spatiotemporal adaptability of the scoring results.

[0017] By adopting the above technical solutions, and by constructing standardized feature vectors and inputting them into the scoring model to perform forward inference, continuous score values ​​reflecting the risk level of grid units can be generated efficiently, thus providing a quantitative basis for subsequent classification. By introducing a dynamic weight adjustment mechanism, the feature contribution can be adjusted according to the completeness of the data source and the adaptability of the scenario, which can improve the robustness of the model to data missingness and regional heterogeneity, thereby enhancing the stability of the scoring and its practical adaptability.

[0018] In one example, this application can be further configured as follows: the fuzzy classification based on the strong convection score of each spatial grid cell and a preset fuzzy membership rule to obtain the strong convection level result of each spatial grid cell specifically includes: Multiple fuzzy membership functions are set, which are constructed using trigonometric functions. The membership intervals are set by adjusting parameters according to historical scoring distribution patterns and business requirements. For each spatial grid cell, the membership degree value under the fuzzy level membership function is calculated for the strong convection score value. The strong convection score values ​​are classified using the maximum membership principle to determine the strong convection level of the corresponding spatial grid cell.

[0019] By adopting the above technical solution, by setting a triangular fuzzy membership function and calculating the membership degree based on the score value, the risk score can be mapped to different severe convective levels in a non-linear manner, thus more closely reflecting the fuzzy evolution characteristics of real weather processes; by using the maximum membership degree principle for classification, the certainty and traceability of the level division can be ensured, thereby facilitating the interpretation of results and the implementation of business.

[0020] The second objective of this invention is achieved through the following technical solution: A strong convection monitoring system based on multi-source data, the strong convection monitoring system based on multi-source data includes: The data acquisition module is used to collect multi-source meteorological data, including radar data, satellite image data, lightning detection data, ground meteorological station data, radiosonde profile data, and numerical weather model products. The data preprocessing module is used to preprocess the multi-source meteorological data to obtain a preprocessed meteorological dataset, and to map the preprocessed meteorological dataset to multiple spatial grid units according to a preset spatial division rule. Each spatial grid unit corresponds to a fixed geographical area and observation time interval. The feature extraction module is used to extract multi-source feature parameters reflecting strong convective activity from the preprocessed meteorological dataset in each spatial grid cell. The multi-source feature parameters include radar reflectivity, cloud top brightness temperature, lightning frequency, CAPE, and wind shear. The scoring reasoning module is used to input the multi-source feature parameters into the pre-trained scoring model for reasoning and analysis, and generate a strong convection score value for each spatial grid cell. The strong convection score value is a continuous real number in the range of 0 to 1, which is used to reflect the risk level of strong convection activity in the corresponding cell and is used for subsequent fuzzy classification level judgment. The classification module is used to perform fuzzy classification based on the strong convection score of each spatial grid cell and the preset fuzzy membership rules to obtain the strong convection level result of each spatial grid cell. The strong convection level classification includes no convection, weak convection, moderate convection and strong convection. The trend analysis module is used to calculate the convection development trend of each spatial grid cell based on the numerical difference between the strong convection score value and the consecutive score values ​​of the previous two moments. The convection development trend includes three states: enhancement, weakening, or stabilization. The layer construction module is used to combine the strong convection level results and convection development trend status of each spatial grid cell to construct a multi-resolution convection monitoring layer, which includes a risk level map, a development trend map, and a confidence map.

[0021] By adopting the above technical solutions, and collecting multi-source meteorological data, it is possible to comprehensively acquire various types of meteorological information reflecting convective activity, thereby enhancing the collaborative perception capability of different meteorological elements. Preprocessing the multi-source meteorological data and mapping it to multiple spatial grid cells unifies the spatiotemporal scale of the data, ensuring the accuracy and consistency of subsequent feature extraction and analysis processes, thus improving the adaptability and operational efficiency of the analysis model. Extracting multi-source feature parameters from each spatial grid cell and inputting them into a scoring model for inference analysis dynamically generates physically meaningful severe convective weather scores, enabling a quantitative assessment of the risk of severe convective weather occurrence. By performing fuzzy classification and trend judgment on the scores, the convective level and its evolution direction can be accurately identified, thereby improving the lead time and reliability of disaster warnings. Constructing multi-resolution convective monitoring layers supports multi-level display and interactive analysis, thereby enhancing the system's practical value in forecasting, early warning, and emergency response scenarios.

[0022] In summary, this application includes the following beneficial technical effects: 1. By collecting multi-source meteorological data, we can comprehensively acquire various types of meteorological information reflecting convective activity, thereby enhancing the ability to collaboratively perceive different meteorological elements; by preprocessing multi-source meteorological data and mapping it to multiple spatial grid units, we can unify the spatiotemporal scale of the data, ensuring the accuracy and consistency of subsequent feature extraction and analysis processes, thereby improving the adaptability and operational efficiency of the analysis model. 2. By extracting multi-source feature parameters from each spatial grid cell and inputting them into the scoring model for inference analysis, a physically meaningful strong convection score value can be dynamically generated, thereby achieving a quantitative assessment of the risk of strong convection. By performing fuzzy classification and development trend judgment on the score value, the convection level and its evolution direction can be accurately identified, thereby improving the lead time and reliability of disaster warnings. By constructing a multi-resolution convection monitoring layer, multi-level display and interactive analysis can be supported, thereby enhancing the practical value of the system in forecasting, early warning, and emergency response scenarios. Attached Figure Description

[0023] Figure 1 This is a flowchart of a strong convection monitoring method based on multi-source data in one embodiment of this application; Figure 2 This is a flowchart illustrating the implementation of step S10 in a strong convection monitoring method based on multi-source data in one embodiment of this application. Figure 3 This is a flowchart illustrating the implementation of step S20 in a strong convection monitoring method based on multi-source data in one embodiment of this application. Figure 4 This is a flowchart illustrating the implementation of step S30 in a strong convection monitoring method based on multi-source data in one embodiment of this application. Figure 5 This is another implementation flowchart of step S40 in a strong convection monitoring method based on multi-source data in one embodiment of this application; Figure 6 This is a flowchart illustrating the implementation of step S40 in a strong convection monitoring method based on multi-source data in one embodiment of this application. Figure 7 This is a flowchart illustrating the implementation of step S50 in a strong convection monitoring method based on multi-source data in one embodiment of this application. Figure 8 This is a principle block diagram of a strong convection monitoring system based on multi-source data in one embodiment of this application. Detailed Implementation

[0024] The present application will be further described in detail below with reference to the accompanying drawings.

[0025] In one embodiment, such as Figure 1 As shown, this application discloses a method for monitoring strong convection based on multi-source data, which specifically includes the following steps: S10: Collects multi-source meteorological data, including radar data, satellite imagery data, lightning detection data, ground meteorological station data, radiosonde profile data, and numerical weather model products.

[0026] Specifically, the system automatically accesses interfaces from multiple heterogeneous meteorological data sources, simultaneously acquiring radar observation data, satellite imagery data, lightning location information, ground weather station records, radiosonde vertical profile data, and numerical weather model output products from a specified time window and geographical area. Radar data is obtained through three-dimensional reflectivity volumetric scanning results collected from regional weather radar stations. Satellite imagery data can be sourced from infrared and visible light channels of GK2A, Himawari-8, and Fengyun series satellites. Lightning detection data is provided by ground-based VLF or LF receivers, providing time series of ground lightning events. Ground station data comes from hourly observation records from a regional automatic weather station network. Radiosonde data obtains temperature and humidity profiles and wind direction and speed structures from radiosonde balloons released periodically from upper-altitude radiosonde stations. Numerical weather model products are based on the analysis field and short-term forecast field output by WRF or ECMWF. For example, in a practical application, the system acquired 12 levels of radar, satellite, and radiosonde data for South China on July 15, 2024, with a 1-hour time resolution, achieving complete spatiotemporal capture of strong convective cloud clusters.

[0027] S20: Preprocess the multi-source meteorological data to obtain the preprocessed meteorological dataset, and map the preprocessed meteorological dataset to multiple spatial grid units according to the preset spatial division rules. Each spatial grid unit corresponds to a fixed geographical area and observation time interval.

[0028] Specifically, the spatial resolution of various meteorological data is uniformly resampled and interpolated. The radar echo grid field and satellite image are unified to a target grid of 0.01°×0.01° according to geographic coordinates. Then, the data of different observation time intervals are resampled and interpolated to be unified to a 10-minute time step. For missing points, spatial neighborhood averaging or temporal linear interpolation is used to fill in the missing points. On this basis, the ground observation data is subjected to pressure and temperature anomaly detection to remove outliers. The radiosonde profile data is hierarchically reconstructed to ensure the integrity of key layers. Finally, a standardized meteorological dataset containing spatial index, time label and multi-source feature fields is generated. For example, in the experiment in North China, the time steps of the original radar echo and FY-4 satellite cloud image were different. After resampling, the two types of data were time aligned, so that cloud top brightness temperature and radar reflectivity can be jointly analyzed at the same time.

[0029] S30: On each spatial grid cell, extract multi-source feature parameters from the preprocessed meteorological dataset to reflect severe convective activity. These multi-source feature parameters include radar reflectivity, cloud top brightness temperature, lightning frequency, CAPE, and wind shear.

[0030] Specifically, for each spatial grid cell, multi-source feature parameters for characterizing convective activity are extracted from the preprocessed meteorological dataset. These include the maximum reflectivity factor extracted from radar reflectivity volume data, the minimum cloud top brightness temperature extracted from satellite infrared brightness temperature data, the number of ground flashes per unit time calculated from lightning detection data, the convective effective potential energy (CAPE) calculated from radiosonde data or model fields, and the 0–6 km vertical wind shear parameters obtained from wind direction and speed profiles. These features constitute a five-dimensional feature vector that is input into the subsequent model. For example, in the early stage of a typical convection event, the CAPE value exceeds 1800 J / kg, the wind shear reaches 20 m / s, and the cloud top brightness temperature drops to -65°C. Based on this, the feature extraction module generates a high-risk feature vector for the corresponding grid and passes it to the scoring model.

[0031] S40: Input the multi-source feature parameters into the pre-trained scoring model for inference analysis to generate a strong convection score value for each spatial grid cell. The strong convection score value is a continuous real number in the range of 0 to 1, which is used to reflect the risk level of strong convection activity in the corresponding cell and is used for subsequent fuzzy classification level judgment.

[0032] Specifically, after standardizing the feature parameters in each spatial grid cell, the data is input into a pre-trained neural network scoring model to perform forward inference. The model outputs a continuous score value between 0 and 1 to characterize the degree of severe convection risk in the current grid cell. To mitigate the impact of missing local data or noise, the input vector dimension weights are dynamically adjusted during the inference process based on the data integrity and feature reliability of the current grid. When some features are missing, compensation is made through historical statistical weights to ensure the stability and spatiotemporal consistency of the output score. For example, in the scenario where sounding data is temporarily missing, the model automatically increases the weights of radar reflectivity and lightning frequency features, so that the scoring results still maintain a sensitive response to the occurrence of severe convection.

[0033] S50: Based on the strong convection score of each spatial grid cell and the preset fuzzy membership rules, fuzzy classification is performed to obtain the strong convection level result of each spatial grid cell. The strong convection level is divided into no convection, weak convection, moderate convection and strong convection.

[0034] Specifically, the score of each grid cell is input into a set of preset triangular fuzzy membership functions. The definition range of the membership functions is distributed between [0,1], corresponding to four levels: no convection, weak convection, moderate convection, and strong convection, respectively. By calculating the membership degree of the score value under each membership function and comparing their maximum values, the final classification level of the grid cell is determined. This fuzzy classification method can effectively avoid abrupt division problems in the critical interval of the score. For example, when a grid cell has a score of 0.67, the model calculates that its membership degree under the "moderate convection" membership function is 0.6 and under the "strong convection" membership function is 0.4. Then the classification result is "moderate convection", achieving physical continuity and classification smoothness.

[0035] S60: Based on the numerical difference between the strong convection score and the consecutive scores of the previous two moments, calculate the convection development trend of each spatial grid cell. The convection development trend includes three states: enhancement, weakening, or stabilization.

[0036] Specifically, the consecutive scores from the previous two time points refer to the strong convection scoring results obtained by the scoring model at two consecutive historical time points before the current observation time, based on inference analysis performed on the current spatial grid cell. To obtain this scoring sequence, it is necessary to first extract the multi-source feature parameters corresponding to the same spatial grid cell at the previous and the time before that time point, and then input them sequentially into the pre-trained scoring model to complete the forward inference operation, obtaining the score value at the previous time point as score(t−1) and the score value at the time before that time point as score(t−2). Using the current time point score(t) as a reference, the scores are calculated by subtracting score(t−1) from score(t−1). The direction and magnitude trend of the change in −score(t−2) form the basis for judging the rate of change and evolution direction of convection intensity within the grid. If the score value shows a continuous upward trend and the increase exceeds the preset enhancement threshold, it is marked as "enhanced" state. If the score value continuously decreases and the decrease exceeds the preset weakening threshold, it is marked as "weakened" state. Otherwise, it is considered as "stable" state. For example, in an actual thunderstorm evolution, if the score values ​​of a certain grid cell at 00:00, 00:05 and 00:10 are 0.35, 0.58 and 0.82 respectively, it indicates that the risk of strong convection is rapidly increasing in a short period of time. This grid is marked as having an "enhanced" development trend, thus providing a criterion for subsequent risk layer overlay and early warning issuance.

[0037] S70: Combining the strong convection level results and convection development trend status of each spatial grid unit, a multi-resolution convection monitoring layer is constructed, which includes a risk level map, a development trend map, and a confidence map.

[0038] Specifically, by combining the severe convection level results and development trend status of each spatial grid unit, a multi-resolution convection monitoring layer is generated. Through spatial coordinate mapping, the layer is displayed on the geographic information platform with color depth to represent level differences, arrows or gradient bars to show trend directions, and transparency or numerical labels to indicate confidence levels. Risk level maps, trend maps, and confidence maps can be overlaid simultaneously on the display, allowing users to view the distribution of severe convection at the regional, city, or grid level. For example, in a rainstorm event in South China, the risk level map highlights high-risk areas with red areas, the trend map shows the attenuation direction with blue arrows, and the confidence map indicates the model's confidence level through a semi-transparent layer, thereby realizing visualized severe convection situation analysis under multi-source data fusion.

[0039] In one embodiment, such as Figure 2 As shown, step S10, which involves collecting multi-source meteorological data, specifically includes: S11: Acquire radar observation data covering the monitoring area. The radar observation data includes three-dimensional reflectivity factor, radial velocity, and echo top height.

[0040] Specifically, the process of acquiring radar observation data includes retrieving volume scan data covering the current monitoring area by calling the distributed data interface of the regional weather radar station, and extracting radar reflectivity factor, radial velocity field and echo top height index obtained from multi-elevation angle scans within a specified time window. Among them, radar reflectivity is used to measure the reflection characteristics of precipitation particles, radial velocity is used to assess wind field structure and shear, and echo top height reflects the vertical depth of convective cloud development. For example, during a rainstorm in the Central Plains region in the summer of 2024, the radar interface received volume scan data from 14 elevation angle layers within 5 minutes, and the cells with echo top heights exceeding 14 kilometers were considered to have potential strong convective risks.

[0041] S12: Acquire multispectral remote sensing image data from geostationary satellites. The image data includes infrared channel brightness temperature, visible light albedo, cloud top brightness temperature, and cloud type identification results.

[0042] Specifically, the process of acquiring satellite image data includes analyzing multispectral remote sensing products from geostationary meteorological satellites. The main approach involves using the brightness temperature data from the infrared channel to identify high-ice cloud structures, obtaining visible light albedo to calculate cloud thickness, extracting cloud top brightness temperature as an approximate indicator of convective top height, and combining this with cloud type classification results obtained from satellite cloud image recognition algorithms to form a multidimensional characterization of cloud cluster properties. For example, during a thunderstorm monitoring event in South China in the autumn of 2023, the infrared and VIS channel images from the FY-4A satellite were used to successfully locate a deep convective cell with a cloud top brightness temperature of -72°C, which assisted in early risk identification.

[0043] S13: Obtain real-time lightning detection data from the lightning location system. The lightning detection data includes ground lightning type, occurrence time, geographical location, and frequency data.

[0044] Specifically, lightning detection data is acquired by establishing a real-time subscription connection with the data service platform of the regional lightning location system. Within the monitoring time range, raw records containing lightning event types, occurrence times, geographical coordinates, and current intensities are received. Events are then categorized and statistically analyzed according to unit time windows to form spatial distribution maps and temporal frequency feature sets. For example, during a severe inland convection event in Shandong, lightning location data recorded 34 lightning events within 5 minutes in a certain grid cell, significantly higher than the surrounding average, providing a strong signal input of sudden electrical activity for the scoring model.

[0045] S14: Collect observation data from ground automatic weather stations, including air temperature, relative humidity, ground wind speed, air pressure, and hourly precipitation.

[0046] Specifically, remote data synchronization channels are connected to ground observation stations to obtain temperature, humidity, wind speed, air pressure, and hourly precipitation data for the most recent hour. Quality control and anomaly detection are performed to ensure timeliness and accuracy. In case of missing data, the missing data is filled by weighted averaging from surrounding stations. For example, during the heavy rainfall event in the middle and lower reaches of the Yangtze River in July 2024, the temperature observation value of a certain county-level grid was 28.7°C, the relative humidity reached 98%, and the wind speed suddenly rose to 6.2 m / s, providing key input for ground disturbance analysis.

[0047] S15: Analyze the vertical profile data provided by the upper-air sounding system. The vertical profile data includes temperature profiles, humidity profiles, wind speed and direction, vertical structure and stability parameters.

[0048] Specifically, standard barometric profiles provided twice daily (00Z / 12Z) by conventional radiosonde stations are used to obtain data on the vertical variation of temperature, humidity, wind direction, and wind speed with altitude. The missing layer areas or the refined structure of the lower layers are reconstructed according to international meteorological standards to ensure the integrity and smoothness of the profile. At the same time, stability indices such as the K-index and the Showalter index are calculated based on the profile to help judge the degree of instability. For example, in a case of afternoon convection in southern China, the radiosonde profile at 12Z showed a significant dry layer between 700hPa and 500hPa, with a CAPE value of 2200J / kg, indicating the potential conditions for strong convection.

[0049] S16: Call the analysis field and forecast products of the numerical weather model. The numerical weather model products include convective available potential energy, wind shear, vertical velocity, isentropic potential temperature and convective parameter field.

[0050] Specifically, by accessing the numerical forecast interface of the regional meteorological data center, the analysis field and hourly forecast products within the specified forecast time and forecast period are obtained. Convective available potential energy (CAPE), 0–6 km wind shear, vertical velocity (ω), isentropic potential temperature (θe), and combined parameters related to convection are extracted and interpolated into the current monitoring grid to form a background field reference for short- to medium-term forecasts. For example, in a case where a strong convection was triggered by a spiral cloud band on the periphery of a typhoon, the WRF model forecast showed that the vertical velocity at the 500 hPa layer in the region reached –0.8 Pa / s, providing a significant weighted reference for the convection score.

[0051] In one embodiment, such as Figure 3 As shown, in step S20, the multi-source meteorological data is preprocessed to obtain a preprocessed meteorological dataset, specifically including: S21: Perform spatial interpolation and grid resampling operations on radar data, satellite image data, and numerical model products from multi-source meteorological data to uniformly map all types of data to a preset spatial grid structure.

[0052] Specifically, unified reconstruction processing is performed on three types of spatially distributed data: radar, satellite, and numerical model data. First, the geographic coordinate system and resolution information corresponding to the original data are read. Then, the grid values ​​of the original data are mapped to a uniformly defined regular grid through bilinear interpolation or inverse distance weighting algorithm. The grid size is set according to business needs, such as a latitude and longitude resolution of 0.01°×0.01° or a UTM projection grid of 1km×1km. When processing numerical model data, height layer interpolation is usually also required to cooperate with subsequent profile extraction tasks. For example, in the field measurement case in the Yangtze River Delta region, the 1km resolution temperature field provided by the WRF model was successfully matched to the standard spatial grid used for convection monitoring through bilinear interpolation, ensuring that data from different sources can be fused and analyzed under the same spatial index structure.

[0053] S22: Perform time alignment processing on various types of observation data. Time alignment processing includes performing time resampling and time interpolation operations on data with different observation time intervals to unify them into a fixed observation time interval.

[0054] Specifically, the timestamps of the original data records are standardized and parsed. A unified observation time benchmark is set according to the sampling frequency differences of various types of data, such as using a uniform interval step size of 10 minutes or 30 minutes. Time interpolation is performed to generate intermediate values ​​for data with low update frequency, such as radiosonde data. For high-frequency data, such as radar or lightning data, resampling or window aggregation is performed to compress to the target time granularity. It is also ensured that various types of data can be correctly correlated and synchronized within the same time step. For example, in a sudden afternoon convection warning in North China, radar volume scan data is refreshed every 6 minutes, lightning data is at a 1-minute granularity, and ground station data is at the 10-minute hourly value. After time alignment, all data are successfully unified to a time frame of 30 minutes, which facilitates subsequent joint input into the scoring model for collaborative inference.

[0055] S23: Use time-series interpolation to complete missing data in lightning detection data and ground meteorological station data from multi-source meteorological data, and perform hierarchical reconstruction of radiosonde profile data in multi-source meteorological data to ensure profile integrity.

[0056] Specifically, interpolation processing is performed on the time-series gaps in the observation point data. For the gaps between lightning data and ground station data, linear interpolation, spline interpolation, or weighted average based on surrounding points are used for reconstruction. At the same time, anomalies are corrected by combining historical trends. For radiosonde profile data, it is necessary to interpolate and fill in the missing pressure layers by interpolation of the upper and lower layers, and reconstruct the vertical structure according to the principle of physical consistency. For example, when there is a data gap in the 500hPa to 400hPa layer of a meteorological station, the missing values ​​are filled by the adjacent time profile and the gradient law of the upper and lower air layers. The reconstructed profile can be continuously input into the stability index calculation to ensure that the model can reliably perceive the vertical structure characteristics. For example, in the monitoring of a tropical depression path in the summer of 2022, the radiosonde data was reconstructed to fill in three missing layers, making the scoring model more accurate in identifying the location of the upper-level dry layer.

[0057] S24: The processed data is formatted and integrated to generate a preprocessed meteorological dataset, which is then mapped to the corresponding spatial grid cells. The preprocessed meteorological dataset includes spatial coordinate indexes, observation time labels, and multi-source feature fields.

[0058] Specifically, the various data after spatial interpolation, temporal alignment, and missing data repair are structurally encapsulated, and a corresponding standardized record format is established for each spatial grid unit. Each record contains a spatial coordinate index, such as grid ID or latitude and longitude coordinates, observation time label, and multi-source feature fields within the corresponding time frame, such as radar maximum reflectivity, cloud top brightness temperature, ground flash frequency, and ground wind speed. This data is saved in a structured form as a dataset that can be directly used in subsequent steps. For data storage, data formats that support multi-dimensional grid structures, such as NetCDF, GeoTIFF, or HDF5, can be used. For example, in the monitoring of the severe thunderstorm process in western Guangdong in 2023, the final preprocessed dataset, with a 30-minute step size and a 1km resolution, covers all radar, satellite, and ground observation features within 4 hours, providing highly consistent input support for model scoring and trend calculation.

[0059] In one embodiment, such as Figure 4 As shown, in step S30, multi-source characteristic parameters reflecting severe convective activity are extracted from the preprocessed meteorological dataset, specifically including: S31: Based on the preset convection physics mechanism and feature extraction rules, the maximum reflectivity, echo top height and vertical gradient of radial velocity at the current moment are extracted from the radar observation data in the preprocessed meteorological dataset in each spatial grid cell.

[0060] Specifically, based on the preset convective physics mechanism and feature extraction rules, a profile analysis is performed on the radar observation data in each spatial grid cell. The maximum radar reflectivity value corresponding to the current observation time and the radial velocity change of each height layer in the vertical direction are extracted and the vertical gradient is calculated accordingly. At the same time, the echo top height in the cell is determined as an indicator representing the volume range of deep convection. In the early stage of strong thunderstorm development, high reflectivity echoes above 50 dBZ can often be observed to rise rapidly to a height of more than 10 km. This combination of indicators can be used to identify potential deep convection development areas.

[0061] S32: Extract cloud top brightness temperature, brightness temperature change rate, and cloud type category index based on brightness temperature from satellite image data in the preprocessed meteorological dataset.

[0062] Specifically, the cloud top brightness temperature value at the current time step is extracted from the satellite image data mapped to the current spatial grid cell, and the rate of change of brightness temperature between two consecutive time steps is calculated to determine the cloud development speed. At the same time, a classification model based on brightness temperature threshold and texture distribution is called to generate cloud type indexes to distinguish different structures such as cumulonimbus, cirrus, or stratus. Among them, if the brightness temperature drops sharply and is accompanied by Cb type cloud type marking, it often indicates an area with strong updraft and active convection.

[0063] S33: Calculate the frequency of ground flashes and the historical cumulative lightning density per unit time from the lightning detection data in the preprocessed meteorological dataset.

[0064] Specifically, the number of ground flashes per unit time is counted from the raw lightning detection data falling within the current grid cell, and historical lightning events over a certain period of time are accumulated to construct lightning density hot zones. A sliding window strategy is used to smooth the frequency distribution to improve feature stability. In contrast to static lightning-free areas, high-density areas with a frequency exceeding 10 times / minute often coincide with the center of strong convective cells, which can help enhance the scoring model's response to electrical activity anomalies.

[0065] S34: Extract near-surface air temperature, humidity and wind speed from the ground weather station data in the preprocessed meteorological dataset, and calculate the ground disturbance intensity by combining the difference between adjacent grids.

[0066] Specifically, near-surface air temperature, relative humidity, and 10-meter wind speed data falling into the current grid cell are extracted from the ground meteorological station data. The parameter difference between the current cell and its four neighboring grids is used as the input feature for calculating the disturbance index. If the sudden change in wind speed is greater than 5 m / s, the air pressure decreases by more than 1.5 hPa, and the temperature and humidity difference between adjacent grid points is significant, it indicates that there is forced uplift or cold pool structure in the local area, suggesting potential uplift triggering conditions or unstable boundary layer structure.

[0067] S35: Extract convective available potential energy, convective suppression energy, wind shear index, and vertical humidity gradient from the radiosonde profile data in the preprocessed meteorological dataset.

[0068] Specifically, the upper-level vertical profile data of the current grid cell is extracted from the preprocessed radiosonde profile data. The CAPE value below 500 hPa is calculated to characterize the lifting potential. The lifting condensation height and the convection inhibition energy (CIN) are combined to determine whether the air mass is easily excited. At the same time, the wind speed difference of 0-6 km is calculated as the wind shear intensity index, and the relative humidity gradient of the 700-850 hPa layer is used to determine the characteristics of water vapor transport in the middle and lower layers. High CAPE, low CIN combined with strong wind shear are typical characteristics of strong convective environment.

[0069] S36: Obtain the vertical velocity, isentropic potential temperature field, and 500 hPa potential temperature anomaly values ​​within the forecast period from the numerical weather model products in the preprocessed meteorological dataset, in order to assist in determining mid-to-high-level instability and uplift triggering conditions.

[0070] Specifically, the output data of numerical model products at the current time point and in the next few hours are selected to extract the vertical velocity field, isentropic potential temperature field and 500 hPa layer potential temperature anomaly value that match the current grid cell space. The isentropic potential temperature rise area is used as a reference for the uplift resistance of the middle warm zone, and the rising area with vertical velocity greater than 0.5 m / s is used as a forced uplift signal. At the same time, the historical monthly mean field is compared to determine whether there are cold vortices, troughs or abnormally low potential temperature characteristics in the 500 hPa layer, in order to help identify potential deep convection triggering mechanisms and environmental support conditions.

[0071] In one embodiment, such as Figure 5 As shown, this method for monitoring strong convection based on multi-source data also includes: S401: Based on historical severe convection samples and corresponding spatiotemporally matched multi-source meteorological feature data, a training sample set is constructed. The training samples include multi-source feature vectors of spatial grid cells and severe convection occurrence labels labeled in real time.

[0072] Specifically, the acquired historical severe convective weather samples include radar images, satellite observation records, and ground station measurement reports from multiple past severe convective events. These samples are matched one-to-one with historical multi-source meteorological data using timestamps and spatial coordinates to generate a training data table with aligned samples. Furthermore, a multi-source feature vector is constructed for each spatial grid cell. The feature dimensions include peak radar reflectivity, cloud top brightness temperature, CAPE, wind shear value, and lightning frequency. At the same time, the presence of severe convective activity in the cell at the corresponding time is labeled as a tag value. For example, historical disaster reports or radar echo intensity continuously exceeding a set threshold can be used as confirmation conditions for the occurrence of severe convection. After the label data is constructed, a training sample set is built in the form of feature vectors and label pairs for subsequent supervised learning model training.

[0073] S402: The model is trained using a multilayer perceptron neural network on the training samples. The model input is a feature vector composed of multi-source feature parameters, and the model output is a continuous strong convection risk score between 0 and 1.

[0074] Specifically, during the training phase, the training sample set is first divided into batches, and a normalization algorithm is used to standardize the feature values ​​of each dimension to make different physical quantities comparable. Then, the standardized multi-source feature vectors are input into the multilayer perceptron neural network model. The model structure includes an input layer, several hidden layers, and an output layer. The hidden layers use the ReLU activation function to enhance nonlinear expression, and the output layer uses the Sigmoid function to compress the output value to the range of 0 to 1 as the estimation result of the severe convection score. The entire training process uses the backpropagation algorithm to optimize the weight parameters to ensure that the model can effectively learn the mapping relationship between different meteorological characteristics and severe convection risk from historical samples.

[0075] S403: During model training, the cross-entropy loss function is used as the optimization objective, and the model overfitting is controlled by early stopping strategy and validation set evaluation.

[0076] Specifically, to avoid the model overfitting to historical data during training, a cross-entropy loss function is introduced as the objective function to measure the difference between the predicted results and the actual labels. At the same time, the ratio of training set to validation set is set, for example, 80% and 20%. After each round of training, the validation set is used to evaluate the change in model performance. When the validation loss does not decrease significantly for several consecutive rounds, an early stopping mechanism is triggered to terminate the training process. In addition, Dropout strategy or L2 regularization constraint can be combined to further suppress model complexity and improve the model's generalization ability in unknown weather scenarios.

[0077] In one embodiment, such as Figure 6As shown, in step S40, the multi-source feature parameters are input into the pre-trained scoring model for inference analysis to generate a strong convection score value for each spatial grid cell, specifically including: S41: In each spatial grid cell, a standardized feature vector is constructed based on multi-source feature parameters.

[0078] Specifically, when performing the feature vector construction operation, the preprocessed multi-source feature parameter values ​​in the current spatial grid cell are first extracted, including radar reflectivity, cloud top brightness temperature, lightning frequency, CAPE value, and wind shear intensity. Then, normalization is performed according to the historical statistical range of the corresponding features, so that all feature values ​​are mapped to a unified range to enhance the convergence performance and generalization ability of the model. For parameters with large discrete variations, such as lightning frequency, logarithmic normalization can be used to suppress the influence of extreme values. For temperature and brightness temperature parameters with obvious upper bounds, Min-Max normalization can be directly used for linear standardization. After the above standardization operation is completed, all features are combined according to the dimension order defined by the model to construct feature vectors and saved to the input tensor matrix of the corresponding grid for model inference.

[0079] S42: Input the standardized feature vector into the pre-trained scoring model, perform forward inference, and output a strong convection score between 0 and 1 to characterize the risk level of strong convection activity in the current grid cell.

[0080] Specifically, in each time-series iteration, all spatial grid cells are traversed sequentially, and the corresponding standardized feature vectors are input into the already trained strong convection scoring model. The forward computation process of the neural network is executed to output a continuous prediction value. This score value is a real number between 0 and 1, representing the probability of a strong convection event occurring in the current grid. The closer the score value is to 1, the higher the convection risk. The model structure can be selected from shallow perceptron networks or lightweight convolutional structures according to actual engineering needs to reduce computational latency and support large-scale parallel processing. In actual deployment, for example, grid cell G_25 in a certain monitoring area outputs a score value of 0.82 after the feature vector is input into the model, which means that the cell has a high probability of strong convection at the current time point. Then, the score value is stored in the grid scoring map of the corresponding time frame for subsequent level classification and trend analysis.

[0081] S43: During the reasoning process, based on the data integrity and scene adaptability of each data source within the current grid cell, a dynamic weight adjustment mechanism is introduced to dynamically adjust the weights of each dimension of the standardized feature vector in order to improve the stability and spatiotemporal adaptability of the scoring results.

[0082] Specifically, a weight adjustment mechanism is introduced before inference execution. The data completeness refers to the proportion of effective observations of each dimension feature in the grid cell. If the proportion is lower than a set threshold, the weight of that feature is penalized by decreasing. The scene adaptability is matched with the similarity between the climate sub-region (such as monsoon region, plateau region, tropical ocean region) to which the region belongs and the representative region in the model training sample. The contribution weight of the corresponding feature is dynamically adjusted according to the degree of matching. The features of each dimension in the standardized feature vector are weighted. If a feature is missing or has a low signal-to-noise ratio in the current grid, its corresponding dimension weight is automatically reduced. Conversely, if a feature has high discriminative ability in the current scene, such as the increased importance of CAPE and ground disturbance index under the background of dry and hot thunderstorms, the model automatically increases the weight coefficient of the relevant feature. Finally, a more reliable and representative strong convection score value is generated, thereby improving the overall scoring mechanism's adaptability to different geographical regions and weather evolution paths.

[0083] Before inputting the standardized feature vectors into the pre-trained scoring model, the data integrity of each data source within the current grid cell is obtained through availability detection of multi-source meteorological data in the spatial and temporal dimensions. Specifically, this includes checking whether the data exists at the current moment, whether there are missing measurements or abnormal delays, whether the data coverage meets the set ratio, and whether the data values ​​exceed reasonable physical ranges. During the detection process, a data availability score is calculated for each data source. For example, radar data has an integrity coefficient calculated based on the proportion of complete pixels in the reflectivity field, while ground meteorological station data has an availability index calculated based on the number of observation points and temporal continuity, thus forming a data integrity matrix for each grid cell. Scene adaptability is then based on historical meteorological model analysis. Based on the analysis and identification results of the current weather situation, by comparing the environmental features (such as CAPE, wind shear, humidity profile, etc.) within the current grid with the feature distribution of historical typical severe convection samples, the sensitivity and discrimination contribution of various observation elements to the formation of severe convection under the current scenario are calculated. For example, the adaptability weight of wind shear features is increased when a front passes, and the weights of CAPE and cloud top brightness temperature features are enhanced in the afternoon thermal convection environment. Finally, the data integrity factor and scenario adaptability factor are obtained through normalization calculation, and used as input references for the dynamic weight adjustment mechanism in the inference stage. This guides the model to automatically adjust the influence weights of each data source during feature fusion, thereby improving the stability of the scoring output and its adaptability to complex weather scenarios.

[0084] In one embodiment, such as Figure 7 As shown, in step S50, fuzzy classification is performed based on the strong convection score of each spatial grid cell and the preset fuzzy membership rules to obtain the strong convection level result for each spatial grid cell, specifically including: S51: Set multiple fuzzy membership functions. The fuzzy membership functions are constructed using trigonometric functions. The membership interval is set by adjusting parameters according to the historical scoring distribution pattern and business requirements.

[0085] Specifically, when constructing fuzzy membership functions, multiple overlapping fuzzy intervals are divided based on the continuous range of strong convection score values. Each interval corresponds to four levels: no convection, weak convection, moderate convection, and strong convection. Triangular membership functions are used as the construction method for membership degrees. The center point, base width, and overlap ratio of each function are jointly determined by the distribution density of historical score samples, empirical statistical laws, and actual business needs. For example, in actual deployment, a score value of 0.2 can be set as the peak value of the "weak convection" membership function, sharing the boundary interval with "no convection" and "moderate convection" to improve the flexibility and fault tolerance of boundary judgment. At the same time, parameters can be adjusted for certain areas with frequent convection to make the membership function better adapt to the regional convection intensity characteristics.

[0086] S52: For each spatial grid cell, calculate the membership degree value under the fuzzy hierarchical membership function for the strong convection score value.

[0087] Specifically, for each spatial grid cell, the continuous strong convection score value output by the current scoring model is obtained and sequentially input into each pre-defined fuzzy membership function to calculate the membership degree, thereby obtaining the corresponding membership degree value under each level of "no convection", "weak convection", "moderate convection" and "strong convection". In practical applications, if a certain grid score value is 0.68, then this value may have a certain membership degree in both the "moderate convection" and "strong convection" levels, for example, 0.4 and 0.6 respectively. Such multiple membership states can better reflect the fuzziness and evolutionary transition of convective activity.

[0088] S53: The strong convection score values ​​are classified according to the maximum membership principle to determine the strong convection level of the corresponding spatial grid cell.

[0089] Specifically, after obtaining the membership values ​​of each level, fuzzy classification logic is executed to determine the final level result. The principle of maximum membership is given priority, that is, the level with the largest membership value is taken as the strong convection level classification result of the current grid cell. At the same time, a weighted decision mechanism or threshold difference condition can be introduced as needed in the boundary scoring interval to avoid frequent jumps. For example, when the membership of "moderate convection" and "strong convection" is very close, the level of the previous moment can be referred as an auxiliary judgment condition, thereby improving the consistency and temporal stability of the level output.

[0090] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0091] In one embodiment, a strong convection monitoring system based on multi-source data is provided, which corresponds one-to-one with the strong convection monitoring method based on multi-source data in the above embodiments. For example... Figure 8 As shown, this strong convection monitoring system based on multi-source data includes a data acquisition module, a data preprocessing module, a feature extraction module, a scoring inference module, a level classification module, a trend analysis module, and a layer construction module. Detailed descriptions of each functional module are as follows: The data acquisition module is used to collect multi-source meteorological data, including radar data, satellite image data, lightning detection data, ground meteorological station data, radiosonde profile data, and numerical weather model products. The data preprocessing module is used to preprocess multi-source meteorological data to obtain a preprocessed meteorological dataset. The preprocessed meteorological dataset is then mapped to multiple spatial grid cells according to a preset spatial division rule. Each spatial grid cell corresponds to a fixed geographical area and observation time interval. The feature extraction module is used to extract multi-source feature parameters reflecting strong convective activity from the preprocessed meteorological dataset in each spatial grid cell. The multi-source feature parameters include radar reflectivity, cloud top brightness temperature, lightning frequency, CAPE, and wind shear. The scoring reasoning module is used to input multi-source feature parameters into a pre-trained scoring model for reasoning and analysis, and generate a strong convection score value for each spatial grid cell. The strong convection score value is a continuous real number in the range of 0 to 1, which is used to reflect the risk level of strong convection activity in the corresponding cell and is used for subsequent fuzzy classification level judgment. The classification module is used to perform fuzzy classification based on the strong convection score of each spatial grid cell and the preset fuzzy membership rules to obtain the strong convection level result of each spatial grid cell. The strong convection level is divided into no convection, weak convection, moderate convection and strong convection. The trend analysis module is used to calculate the convection development trend of each spatial grid cell based on the numerical difference between the strong convection score and the consecutive score values ​​of the previous two moments. The convection development trend includes three states: enhancement, weakening, or stabilization. The layer construction module is used to combine the strong convection level results and convection development trend status of each spatial grid cell to construct a multi-resolution convection monitoring layer, which includes a risk level map, a development trend map, and a confidence map.

[0092] Optionally, the data acquisition module includes: The radar observation and acquisition submodule is used to acquire radar observation data covering the monitoring area. The radar observation data includes three-dimensional reflectivity factor, radial velocity and echo top height. The satellite image acquisition submodule is used to acquire multispectral remote sensing image data from geostationary satellites. The image data includes infrared channel brightness temperature, visible light albedo, cloud top brightness temperature, and cloud type identification results. The lightning detection and acquisition submodule is used to acquire real-time lightning detection data from the lightning location system. The lightning detection data includes ground lightning type, occurrence time, geographical location, and frequency data. The ground observation and acquisition submodule is used to collect observation data from ground automatic weather stations. The observation data includes air temperature, relative humidity, ground wind speed, air pressure and hourly precipitation. The sounding data parsing submodule is used to parse the vertical profile data provided by the upper-air sounding system. The vertical profile data includes temperature profiles, humidity profiles, wind speed and direction, vertical structure and stability parameters. The numerical model product invocation submodule is used to invoke the analysis field and forecast products of the numerical weather model, which include convective available potential energy, wind shear, vertical velocity, isentropic potential temperature and convective parameter field.

[0093] Optionally, the data preprocessing module includes: The Spatial Unification Submodule is used to perform spatial interpolation and grid resampling operations on radar data, satellite image data and numerical model products in multi-source meteorological data, so that all types of data are uniformly mapped to the preset spatial grid structure. The time alignment submodule is used to perform time alignment processing on various types of observation data. Time alignment processing includes performing time resampling and time interpolation operations on data with different observation time intervals to unify them into a fixed observation time interval. The data completion and structure reconstruction submodule is used to complete missing data in lightning detection data and ground meteorological station data in multi-source meteorological data by using time-series interpolation, and to perform hierarchical reconstruction of radiosonde profile data in multi-source meteorological data to ensure profile integrity. The standard data construction submodule is used to format and integrate the processed data to generate a preprocessed meteorological dataset, which is then mapped to the corresponding spatial grid cells. The preprocessed meteorological dataset includes spatial coordinate indexes, observation time labels, and multi-source feature fields.

[0094] Optionally, the feature extraction module includes: The radar feature extraction submodule is used to extract the maximum reflectivity, echo top height and vertical gradient of radial velocity at the current moment from the radar observation data in the preprocessed meteorological dataset in each spatial grid cell, based on the preset convection physics mechanism and feature extraction rules. The satellite feature extraction submodule is used to extract cloud top brightness temperature, brightness temperature change rate, and cloud type category index based on brightness temperature from satellite image data in the preprocessed meteorological dataset. The lightning feature extraction submodule is used to calculate the frequency of ground flashes and the historical cumulative lightning density per unit time from the lightning detection data in the preprocessed meteorological dataset. The ground feature extraction submodule is used to extract near-surface air temperature, humidity and wind speed from ground meteorological station data in the preprocessed meteorological dataset, and calculate the ground disturbance intensity by combining the difference between adjacent grids. The radiosonde feature extraction submodule is used to extract convective available potential energy, convective suppression energy, wind shear index and vertical humidity gradient from the radiosonde profile data in the preprocessed meteorological dataset. The pattern feature extraction submodule is used to obtain the vertical velocity, isentropic potential temperature field and 500hPa layer potential temperature anomaly value within the forecast period from the numerical weather model products in the preprocessed meteorological dataset, in order to help determine the instability and uplift triggering conditions in the middle and upper layers.

[0095] Optionally, this strong convection monitoring system based on multi-source data also includes: The training sample construction module is used to construct a training sample set based on historical strong convection real-time samples and corresponding spatiotemporally matched multi-source meteorological feature data. The training samples include multi-source feature vectors of spatial grid cells and strong convection occurrence labels labeled in real-time. The model training module is used to train the model on the training samples using a multilayer perceptron neural network. The model input is a feature vector composed of multi-source feature parameters, and the model output is a continuous strong convection risk score value between 0 and 1. The training optimization control module is used to use the cross-entropy loss function as the optimization objective during model training, and to control model overfitting by using an early stopping strategy and validation set evaluation.

[0096] Optionally, the scoring reasoning module includes: The feature vector construction submodule is used to construct a standardized feature vector based on multi-source feature parameters in each spatial grid cell; The model inference execution submodule is used to input standardized feature vectors into a pre-trained scoring model, perform forward inference operations, and output a strong convection score value between 0 and 1 to characterize the risk level of strong convection activity in the current grid cell. The dynamic weight adjustment submodule is used to introduce a dynamic weight adjustment mechanism during the inference process, based on the data integrity and scenario adaptability of each data source within the current grid cell. This mechanism dynamically adjusts the weights of each dimension of the standardized feature vector to improve the stability and spatiotemporal adaptability of the scoring results.

[0097] Optional, the rating classification module includes: The membership function setting submodule is used to set multiple fuzzy membership functions. The fuzzy membership functions are constructed using trigonometric functions, and the membership intervals are set by adjusting parameters according to historical scoring distribution patterns and business requirements. The membership calculation submodule is used to calculate the membership value under the fuzzy hierarchical membership function for each spatial grid cell's strong convection score value. The fuzzy classification submodule is used to classify the strong convection score values ​​using the maximum membership principle and determine the strong convection level result to which the corresponding spatial grid cell belongs.

[0098] For specific limitations regarding a strong convection monitoring system based on multi-source data, please refer to the limitations of a strong convection monitoring method based on multi-source data mentioned above, which will not be repeated here. Each module in the aforementioned strong convection monitoring system based on multi-source data can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0099] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above.

[0100] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for monitoring strong convection based on multi-source data, characterized in that, The method for monitoring strong convection based on multi-source data includes: Collect multi-source meteorological data, including radar data, satellite image data, lightning detection data, ground meteorological station data, radiosonde profile data, and numerical weather model products; The multi-source meteorological data is preprocessed to obtain a preprocessed meteorological dataset, and the preprocessed meteorological dataset is mapped to multiple spatial grid units according to a preset spatial division rule. Each spatial grid unit corresponds to a fixed geographical area and observation time interval. In each spatial grid cell, multi-source feature parameters reflecting severe convective activity are extracted from the preprocessed meteorological dataset. These multi-source feature parameters include radar reflectivity, cloud top brightness temperature, lightning frequency, CAPE, and wind shear. The multi-source feature parameters are input into a pre-trained scoring model for inference and analysis to generate a strong convection score value for each spatial grid cell. The strong convection score value is a continuous real number in the range of 0 to 1, which is used to reflect the risk level of strong convection activity in the corresponding cell and is used for subsequent fuzzy classification level judgment. Fuzzy classification is performed based on the strong convection score of each spatial grid cell and the preset fuzzy membership rules to obtain the strong convection level result of each spatial grid cell. The strong convection level classification includes no convection, weak convection, moderate convection and strong convection. Based on the numerical difference between the strong convection score and the consecutive scores of the previous two moments, the convection development trend of each spatial grid cell is calculated, and the convection development trend includes three states: enhancement, weakening or stabilization. By combining the strong convection level results and convection development trend status of each spatial grid cell, a multi-resolution convection monitoring layer is constructed, which includes a risk level map, a development trend map, and a confidence level map.

2. The method for monitoring strong convection based on multi-source data according to claim 1, characterized in that, The collection of multi-source meteorological data specifically includes: Acquire radar observation data covering the monitoring area, the radar observation data including three-dimensional reflectivity factor, radial velocity and echo top height; Acquire multispectral remote sensing image data from geostationary satellites, including infrared channel brightness temperature, visible light albedo, cloud top brightness temperature, and cloud type identification results. Acquire real-time lightning detection data from the lightning location system, including ground lightning type, occurrence time, geographical location, and frequency data; Collect observation data from ground-based automatic weather stations, including air temperature, relative humidity, surface wind speed, air pressure, and hourly precipitation. Analyze the vertical profile data provided by the upper-air sounding system, which includes temperature profile, humidity profile, wind speed and direction, vertical structure and stability parameters; The analysis field and forecast products of the numerical weather model are invoked, including convective available potential energy, wind shear, vertical velocity, isentropic potential temperature, and convective parameter fields.

3. The method for monitoring strong convection based on multi-source data according to claim 1, characterized in that, The preprocessing of the multi-source meteorological data to obtain a preprocessed meteorological dataset specifically includes: Spatial interpolation and grid resampling operations are performed on radar data, satellite image data and numerical model products in the multi-source meteorological data to uniformly map all types of data to a preset spatial grid structure. Time alignment processing is performed on various types of observation data. The time alignment processing includes performing time resampling and time interpolation operations on data with different observation time intervals to unify them into a fixed observation time interval. Time-series interpolation is used to complete missing data in lightning detection data and ground meteorological station data in the multi-source meteorological data, and hierarchical reconstruction is performed on radiosonde profile data in the multi-source meteorological data to ensure profile integrity. The processed data is formatted and integrated to generate the preprocessed meteorological dataset, which is then mapped to the corresponding spatial grid cells. The preprocessed meteorological dataset includes spatial coordinate indexes, observation time labels, and multi-source feature fields.

4. The method for monitoring strong convection based on multi-source data according to claim 1, characterized in that, The extraction of multi-source feature parameters reflecting severe convective activity from the preprocessed meteorological dataset specifically includes: Based on the preset convection physics mechanism and feature extraction rules, the maximum reflectivity, echo top height and vertical gradient of radial velocity at the current moment are extracted from the radar observation data in the preprocessed meteorological dataset in each spatial grid cell. Cloud top brightness temperature, brightness temperature change rate, and cloud type category index based on brightness temperature determination are extracted from the satellite image data in the preprocessed meteorological dataset. The frequency of ground flashes and the historical cumulative lightning density per unit time are calculated from the lightning detection data in the preprocessed meteorological dataset. Near-surface air temperature, humidity and wind speed are extracted from the ground weather station data in the preprocessed meteorological dataset, and the ground disturbance intensity is calculated by combining the difference between adjacent grids. The convective potential energy, convective suppression energy, wind shear index, and vertical humidity gradient are extracted from the radiosonde profile data in the preprocessed meteorological dataset. The vertical velocity, isentropic potential temperature field, and 500 hPa layer potential temperature anomaly values ​​within the forecast period are obtained from the numerical weather model products in the preprocessed meteorological dataset to assist in determining mid-to-high-level instability and uplift triggering conditions.

5. The method for monitoring strong convection based on multi-source data according to claim 1, characterized in that, The method for monitoring strong convection based on multi-source data also includes: A training sample set is constructed based on historical severe convection samples and corresponding spatiotemporally matched multi-source meteorological feature data. The training samples include multi-source feature vectors of spatial grid cells and severe convection occurrence labels labeled with real-time data. The training samples are trained using a multilayer perceptron neural network. The model input is a feature vector composed of multi-source feature parameters, and the model output is a continuous strong convection risk score value between 0 and 1. During model training, the cross-entropy loss function is used as the optimization objective, and the model overfitting is controlled by early stopping strategy and validation set evaluation. The trained model is used as the pre-trained scoring model.

6. The method for monitoring strong convection based on multi-source data according to claim 1, characterized in that, The step of inputting the multi-source feature parameters into a pre-trained scoring model for inference analysis to generate a strong convection score value for each spatial grid cell specifically includes: In each spatial grid cell, a standardized feature vector is constructed based on the multi-source feature parameters; The standardized feature vector is input into the pre-trained scoring model, and forward inference is performed to output a strong convection score value between 0 and 1, which is used to characterize the risk level of strong convection activity in the current grid cell. During the reasoning process, a dynamic weight adjustment mechanism is introduced based on the data integrity and scene adaptability of each data source within the current grid cell. This mechanism dynamically adjusts the weights of each dimension of the standardized feature vector to improve the stability and spatiotemporal adaptability of the scoring results.

7. The method for monitoring strong convection based on multi-source data according to claim 1, characterized in that, The process of performing fuzzy classification based on the strong convection score of each spatial grid cell and a preset fuzzy membership rule to obtain the strong convection level result for each spatial grid cell specifically includes: Multiple fuzzy membership functions are set, which are constructed using trigonometric functions. The membership intervals are set by adjusting parameters according to historical scoring distribution patterns and business requirements. For each spatial grid cell, the membership degree value under the fuzzy level membership function is calculated for the strong convection score value. The strong convection score values ​​are classified using the maximum membership principle to determine the strong convection level of the corresponding spatial grid cell.

8. A strong convection monitoring system based on multi-source data, characterized in that, The strong convection monitoring system based on multi-source data includes: The data acquisition module is used to collect multi-source meteorological data, including radar data, satellite image data, lightning detection data, ground meteorological station data, radiosonde profile data, and numerical weather model products. The data preprocessing module is used to preprocess the multi-source meteorological data to obtain a preprocessed meteorological dataset, and to map the preprocessed meteorological dataset to multiple spatial grid units according to a preset spatial division rule. Each spatial grid unit corresponds to a fixed geographical area and observation time interval. The feature extraction module is used to extract multi-source feature parameters reflecting strong convective activity from the preprocessed meteorological dataset in each spatial grid cell. The multi-source feature parameters include radar reflectivity, cloud top brightness temperature, lightning frequency, CAPE, and wind shear. The scoring reasoning module is used to input the multi-source feature parameters into the pre-trained scoring model for reasoning and analysis, and generate a strong convection score value for each spatial grid cell. The strong convection score value is a continuous real number in the range of 0 to 1, which is used to reflect the risk level of strong convection activity in the corresponding cell and is used for subsequent fuzzy classification level judgment. The classification module is used to perform fuzzy classification based on the strong convection score of each spatial grid cell and the preset fuzzy membership rules to obtain the strong convection level result of each spatial grid cell. The strong convection level classification includes no convection, weak convection, moderate convection and strong convection. The trend analysis module is used to calculate the convection development trend of each spatial grid cell based on the numerical difference between the strong convection score value and the consecutive score values ​​of the previous two moments. The convection development trend includes three states: enhancement, weakening, or stabilization. The layer construction module is used to combine the strong convection level results and convection development trend status of each spatial grid cell to construct a multi-resolution convection monitoring layer, which includes a risk level map, a development trend map, and a confidence map.

Images