A method for diagnosing a full lightning activity region for a lightning detection data missing area

Abstract

The application discloses a kind of full lightning activity area diagnostic methods for lightning detection data missing area, including collecting lightning activity data and carrying out space-time consistency matching, carry out convection cloud area and lightning identification, construct lightning label set according to lightning identification result in convection cloud area, carry out feature calculation and obtain multi-dimensional infrared brightness temperature feature, carry out feature screening and obtain multi-dimensional infrared brightness temperature representative feature, according to lightning label set and corresponding multi-dimensional infrared brightness temperature representative feature carry out multi-model integration training and optimization, carry out model evaluation and select comprehensive optimal lightning diagnostic model, input comprehensive optimal lightning diagnostic model and obtain lightning diagnostic result with the multi-dimensional infrared channel brightness temperature data to be diagnosed.It can realize the identification of lightning activity area kilometer level, solve the technical bottleneck of efficient monitoring of lightning activity in remote areas, provide an implementation means for lightning monitoring and early warning in remote areas, thereby effectively supporting the lightning protection and disaster reduction work in the corresponding area.

Description

Technical Field

[0001] This invention relates to the field of severe convective weather monitoring and early warning technology, and in particular to a method for diagnosing full-flash activity areas in regions lacking lightning detection data. Background Technology

[0002] Lightning activity is a significant component of severe convective weather, possessing immense destructive power and posing a serious threat to people's lives and property, aerospace, power transmission, and outdoor operations. Timely and accurate monitoring and early warning of lightning activity is a core task in the field of meteorological disaster prevention and mitigation, and has important social and economic significance.

[0003] Currently, mainstream lightning monitoring methods mainly rely on ground-based lightning location networks and weather radar systems. These technologies are widely used in cities and areas with well-developed infrastructure. However, in remote areas, complex geographical environments, difficulties in power supply and communication, and high maintenance costs result in low deployment density of these devices, leading to missing or low-quality lightning detection data. This restricts the improvement of lightning monitoring and early warning capabilities and meets the growing industry safety needs. Geostationary meteorological satellites have unique advantages such as wide coverage, high spatiotemporal resolution, and continuous observation. Their multispectral infrared observation data are closely related to convective cloud development and cloud top microphysical processes, providing a new data source to compensate for the shortcomings of ground-based observations. In recent years, with the development of machine learning technology, intelligent identification of meteorological elements using satellite data has become a research hotspot. Therefore, this invention proposes a diagnostic method for all-flash activity areas in regions lacking lightning detection data. It innovatively integrates multidimensional infrared channel brightness temperature data from geostationary meteorological satellites with high-precision three-dimensional all-flash detection data of the region. A diagnostic model is constructed using advanced machine learning algorithms, and the model is then applied to regions lacking lightning detection data. This enables high-precision, kilometer-level diagnostic monitoring of all-flash activity in these areas, which has significant scientific value and application prospects. Summary of the Invention

[0004] The purpose of this invention is to provide a method for diagnosing full-flash activity areas in regions where lightning detection data is lacking.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution:

[0006] This invention includes the following steps:

[0007] Collect lightning activity data and perform spatiotemporal consistency matching; the lightning activity data includes multidimensional infrared channel brightness temperature data and lightning location data;

[0008] Based on the multidimensional infrared channel brightness temperature data, convective cloud regions are identified, and lightning is identified from the lightning location data. A lightning tag set is constructed in the convective cloud regions based on the lightning identification results. The lightning tag set includes a positive sample set and a negative sample set.

[0009] Multidimensional brightness temperature features are obtained by performing feature calculations based on the multidimensional infrared channel brightness temperature data, and multidimensional brightness temperature representative features are obtained by filtering the multidimensional brightness temperature feature set based on the lightning tag.

[0010] Based on the lightning tag set and corresponding multidimensional brightness temperature representative features, multi-model integration training and optimization are performed. Model evaluation is conducted to select the comprehensive optimal lightning diagnosis model. The lightning activity data to be diagnosed is input into the comprehensive optimal lightning diagnosis model to obtain the lightning diagnosis result.

[0011] Furthermore, the method for constructing the lightning tag set includes:

[0012] Multidimensional infrared brightness temperature data were collected by a geostationary meteorological satellite, and lightning location data were obtained by a regional three-dimensional full-flash detection system; the wavelength range of each channel of the multidimensional infrared brightness temperature data is 3.9μm~13.3μm.

[0013] Cloud regions with brightness temperature values ​​less than the brightness temperature threshold within a fixed wavelength channel are defined as convective cloud regions.

[0014] The minimum convex hull of lightning is calculated using a lightning radiation source attribution algorithm. Within the convective cloud region, satellite grids covered or traversed by the convex hull are marked as areas with lightning activity, while the remaining satellite grids are marked as areas without lightning activity. Satellite grids with lightning activity are used as positive samples, and satellite grids without lightning activity are used as negative samples to form a lightning label set. The minimum convex hull of lightning is used to characterize the range of lightning activity.

[0015] Furthermore, the method for obtaining multidimensional brightness temperature features includes:

[0016] Multidimensional infrared channel brightness temperature data is used as the original brightness temperature feature;

[0017] Brightness temperature difference characteristics are obtained by calculating the brightness temperature difference between different infrared channels.

[0018] The horizontal gradient features are obtained by calculating the horizontal spatial gradient of the brightness temperature data and brightness temperature difference of each channel. The horizontal gradient features are based on satellite grids, and the absolute value of the average spatial change rate of the original brightness temperature features or brightness temperature difference features between the satellite grid and its four neighboring satellite grids is taken as the horizontal gradient feature of the corresponding satellite grid. The number of horizontal gradient features is equal to the sum of the number of original brightness temperature features and the number of brightness temperature difference features.

[0019] The multidimensional brightness temperature feature consists of the original brightness temperature feature, the brightness temperature difference feature, and the horizontal gradient brightness temperature feature.

[0020] Furthermore, the method for obtaining multidimensional brightness temperature representative features includes:

[0021] The correlation between multidimensional brightness temperature features and lightning tag sets is calculated to obtain the lightning correlation degree of each multidimensional brightness temperature feature. The mutual redundancy among multidimensional brightness temperature features is calculated to obtain the feature redundancy degree of each multidimensional brightness temperature feature. The feature importance of each multidimensional brightness temperature feature is determined based on the lightning correlation degree and the lightning redundancy degree. The multidimensional brightness temperature features are sorted and selected according to the feature importance to obtain representative multidimensional brightness temperature features. The feature importance is positively correlated with the lightning correlation degree and negatively correlated with the feature redundancy degree.

[0022] Furthermore, the method for multi-model ensemble training and optimization includes:

[0023] Multidimensional brightness temperature representative features and corresponding lightning labels are used as a set of lightning analysis data. Lightning analysis data at all times are collected to form a lightning comprehensive set. The lightning comprehensive set is randomly divided into a training set and a test set in an 8:2 ratio. The training set is used to train the lightning activity diagnostic base model.

[0024] The base models for lightning activity diagnostics include lightweight gradient boosting machines, gradient boosting decision trees, random forests, logistic regression, extreme gradient boosting, and artificial neural networks;

[0025] Based on experience and theoretical analysis, the reasonable search range of key hyperparameters of various lightning activity diagnostic base models is analyzed. A multidimensional parameter grid space is constructed and an exhaustive search is performed. Combined with 5-fold cross-validation, the hyperparameters of the lightning activity diagnostic base models are systematically optimized.

[0026] Furthermore, the method for selecting the comprehensive optimal lightning diagnostic model includes:

[0027] A multi-dimensional evaluation index system is constructed. The performance of the lightning activity diagnostic base model is evaluated using training and test sets to obtain multi-dimensional evaluation indexes. Based on the multi-dimensional evaluation indexes, a comprehensive evaluation is conducted to determine the optimal lightning diagnostic model. The multi-dimensional evaluation indexes include accuracy, precision, recall, and F1 score.

[0028] The generalization ability of the comprehensive optimal lightning diagnostic model was verified using actual weather events;

[0029] The multidimensional infrared channel brightness temperature data is input into the comprehensive optimal lightning diagnosis model to obtain the lightning diagnosis results.

[0030] The beneficial effects of this invention are:

[0031] This invention is a diagnostic method for all-flash activity areas in regions with missing lightning detection data. Compared with existing technologies, this invention has the following technical advantages:

[0032] This invention reduces reliance on human experience and significantly improves the efficiency and modernization of lightning monitoring through spatiotemporal consistency matching, convective cloud region and lightning identification, feature calculation and screening, model training and evaluation steps. Combined with geostationary meteorological satellites and regional three-dimensional full-flash detection systems, it can effectively cover blind spots in traditional lightning monitoring, thus solving the problem of missing lightning monitoring data. By selecting the most discriminative core feature subset, it effectively improves model efficiency and reduces the risk of overfitting, ensuring high accuracy and stability of diagnostic results. The model has strong generalization ability and can be transferred to data-scarce areas with similar climatic backgrounds, meeting the needs of multiple industries for refined and timely lightning risk early warning, and providing key data support for their safe production and dispatching decisions. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the steps of a method for diagnosing total flash activity areas in regions with missing lightning detection data, as described in this invention.

[0034] Figure 2 This is a schematic diagram of lightning activity space tag allocation in an embodiment of the present invention;

[0035] Figure 3 This is a comparative analysis of ROC-AUC and AP of the lightning activity diagnostic base model on the test set in an embodiment of the present invention;

[0036] Figure 4 The results show the application effects of the optimal model in different actual weather processes in the embodiments of the present invention. Detailed Implementation

[0037] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0038] The present invention provides a method for diagnosing total flash activity areas in regions with missing lightning detection data, comprising the following steps:

[0039] Taking the construction and verification of a three-dimensional total flash monitoring and early warning model for the South China Sea based on geostationary meteorological satellites as an example, this embodiment constructs and verifies a three-dimensional total flash monitoring model applicable to the vast ocean area of ​​the XX marine region using three-dimensional total flash data of XX islands and infrared channel data of YY geostationary meteorological satellites. Figure 1 As shown, this embodiment includes the following steps:

[0040] Collect lightning activity data and perform spatiotemporal consistency matching; the lightning activity data includes multidimensional infrared channel brightness temperature data and lightning location data;

[0041] Based on the multidimensional infrared channel brightness temperature data, convective cloud regions are identified, and lightning is identified from the lightning location data. A lightning tag set is constructed in the convective cloud regions based on the lightning identification results. The lightning tag set includes a positive sample set and a negative sample set.

[0042] Multidimensional brightness temperature features are obtained by performing feature calculations based on the multidimensional infrared channel brightness temperature data, and multidimensional brightness temperature representative features are obtained by filtering the multidimensional brightness temperature feature set based on the lightning tag.

[0043] Based on the lightning tag set and corresponding multidimensional brightness temperature representative features, multi-model integration training and optimization are performed. Model evaluation is conducted to select the comprehensive optimal lightning diagnosis model. Multidimensional infrared channel brightness temperature data is input into the comprehensive optimal lightning diagnosis model to obtain lightning diagnosis results.

[0044] In the actual assessment, the first step was to acquire lightning location data in the XX marine area: the low-frequency lightning detection array (HN-LFEDA) deployed by Organization A in the XX marine area was used to acquire three-dimensional total lightning observation data. The system contains 15 detection substations with a detection frequency range of 160 Hz to 600 kHz. The lightning discharge process was finely located based on the time difference of arrival method. The range for acquiring three-dimensional total lightning data was a rectangular area with a length and width of 100 km (latitude and longitude range of 109.94-110.94 °E, 19.40-20.40 °N) centered on the center of the station network, to ensure the optimal detection efficiency of the LFEDA system.

[0045] The second step is to acquire multi-dimensional infrared brightness temperature data for the XX region: using L1 level brightness temperature data from 10 infrared channels of the Advanced Imager (AHI) of the YY geostationary meteorological satellite, with a wavelength coverage range of 3.9-13.3 μm and units of Kelvin (K). The data time resolution is 30 minutes (actually, observations were taken every 10 minutes, but to save data resources, one data point every 30 minutes was selected), and the spatial resolution is 2 km. The latitude and longitude are cropped to obtain rectangular areas containing lightning data and infrared channel data for the South China Sea region.

[0046] The third step is to construct a lightning tag set: Select HN-LFEDA lightning detection data within the corresponding satellite observation time (10 minutes for the YY geostationary orbit meteorological satellite). For each lightning event, calculate the minimum convex hull area encompassing all its radiation sources as the spatial impact range of that lightning event. Simultaneously, use an infrared channel brightness temperature value of less than 238 K at a wavelength of 10.4 μm as the criterion for cloud regions to identify potential convective activity areas. Mark satellite grids covered or passed by the lightning convex hull as "1" (lightning present), and mark the remaining grids within the convective activity area as "0" (no lightning). Figure 2 As shown (where the black area is marked as a lightning grid and the other areas are non-lightning grids), when the lightning convex shell extends beyond the boundary of the rectangular area, only the part within the area is included in the marked range;

[0047] The fourth step is to perform feature calculations to obtain multidimensional brightness temperature features:

[0048] (1) Determine the original brightness temperature characteristics. The brightness temperature values ​​of 10 infrared channels (BT07-16) are used directly as the basic characteristics. The specific channels and their physical meanings are as follows: BT07 / 3.9μm / shortwave infrared channel is used for the identification of low clouds and fog; BT08 / 6.2μm / upper tropospheric water vapor channel; BT09 / 6.9μm / middle tropospheric water vapor channel; BT10 / 7.3μm / lower tropospheric water vapor channel; BT11 / 8.6μm / cloud phase identification channel; BT12 / 9.6μm / ozone absorption channel; BT13 / 10.4μm / infrared window channel is used for cloud top and surface temperature estimation; BT14 / 11.2μm / infrared window channel; BT15 / 12.4μm / infrared window channel; BT16 / 13.3μm / CO2 absorption channel is used for cloud top height estimation.

[0049] (2) Calculate the brightness temperature difference characteristics. Based on the physical correlation between different infrared channels, calculate the cloud microphysical property difference: BT14_07 (representing BT14-BT07, the same below) / reflects the intensity of convective cloud development, BT11_14 / used for cloud phase identification (positive value indicates the existence of ice cloud), BT14_15 / reflects cloud particle size information BT(11_14)_(14_15) / three-spectral difference (suitable for convective initiation identification); calculate cloud top height difference: BT13_15 / cloud top height estimation (the larger the difference, the higher the cloud top), BT16_14 / cloud top height relative to the tropopause, BT08_14 / reflects whether the cloud top penetrates the middle and upper water vapor layer; calculate water vapor structure difference: BT08_10 / difference between upper and lower water vapor, BT08_13 / upper water vapor content indication, BT10_14 / lower water vapor content indication;

[0050] (3) Calculate the horizontal gradient features: For each original brightness temperature channel and brightness temperature difference, calculate its horizontal gradient, denoted as GBT. The gradient calculation uses the average of the spatial change rate between each grid pixel and its four surrounding grid pixels, and uses the absolute value of the brightness temperature horizontal gradient to obtain 20 horizontal gradient feature parameters.

[0051] Step 5, Feature Optimization and Screening: Based on the existing Minimum Redundancy Maximum Relevance (mRMR) algorithm, the 40 constructed features are screened. Based on mutual information theory, the correlation between the multidimensional brightness temperature features and the lightning tag set is calculated to obtain the lightning correlation of each multidimensional brightness temperature feature. The mutual redundancy between the multidimensional brightness temperature features is calculated to obtain the feature redundancy of each multidimensional brightness temperature feature. The feature importance of each multidimensional brightness temperature feature is determined, sorted, and screened to obtain representative multidimensional brightness temperature features. The specific implementation method is as follows: In Python, the mrmr_classif function is called to process the 40-dimensional feature vector. The input parameters include the multidimensional brightness temperature feature matrix and the lightning tag, and the top 10 core feature variable subsets in importance ranking are obtained (BT09, BT10, BT11, BT12, BT13, BT13_15, BT14_15, BT16_14, GBT08, GBT14_07).

[0052] Step 5, multi-model integration training and optimization: a random undersampling strategy is used to sample negative samples to balance the number of positive and negative samples (due to the relative scarcity of lightning events, there is a significant imbalance between positive and negative samples). The multidimensional brightness temperature representative features and the corresponding lightning labels are used as a set of lightning analysis data. Lightning analysis data at all times are collected to form a lightning comprehensive set. The lightning comprehensive set is randomly divided into a training set and a test set in an 8:2 ratio.

[0053] The training set is used to train the base model for lightning activity diagnosis (lightweight gradient booster, gradient boosting decision tree, random forest, logistic regression, extreme gradient boosting and artificial neural network). The six machine learning classifiers can be called in Python using the LGBMClassifier, GradientBoostingClassifier, RandomForestClassifier, XGBClassifier and LogisticRegression libraries respectively. The artificial neural network architecture is built using PyTorch.

[0054] Based on empirical and theoretical analysis, a reasonable search range for key hyperparameters of various lightning activity prediction base models was determined. A multi-dimensional parameter grid space was constructed, and an exhaustive search was performed within the predefined parameter grid space to evaluate the impact of each parameter combination on model performance. Systematic hyperparameter optimization of each model was then performed using 5-fold cross-validation. Specifically, the training dataset was randomly divided into five equal subsets. Four subsets were used for model training, and the remaining subset was used for validation. This process was repeated five times to ensure that each subset served as the validation set, thus obtaining an unbiased estimate of model performance. Accuracy was used as the primary optimization objective, and the optimal hyperparameter configurations for each algorithm were determined. The specific parameter settings are shown in Table 1.

[0055] Table 1. Parameters of each lightning activity diagnostic baseline model

[0056]

[0057] Step 6, Multi-model Integration Training and Optimization: For the six trained models, calculate the evaluation metric (accuracy) on both the training and test sets. Precision = Recall = F1 Score = Where TP represents true positives (the number of samples correctly classified as having lightning), TN represents true negatives (the number of samples correctly classified as not having lightning), FP represents false positives (the number of samples incorrectly predicted as having lightning), and FN represents false negatives (the number of samples incorrectly predicted as not having lightning). ROC-AUC is the area under the recall and false positive curves, and AP is the area under the precision-recall curve. The evaluation results based on the training and test sets are detailed in Table 2.

[0058] Table 2 Evaluation results based on training and test sets

[0059]

[0060] As can be seen, XGBoost outperforms other models in all metrics on both the training and test sets;

[0061] Figure 3The evaluation results are the area under the recall and false positive rate curves (ROC-AUC) and the area under the precision-recall curve (AP) of each model on the test set. ROC-AUC measures the classifier's discriminative ability by plotting the relationship between recall and false positive rate under different classification thresholds. Its area under the curve is a quantitative indicator of the overall performance of the model, with a value between 0 and 1. The closer the value is to 1, the better the model performance. AP is used to evaluate the classifier performance on imbalanced datasets by plotting the relationship between the model's precision and recall under different classification thresholds. The closer its area under the curve is to 1, the more the model can maintain high precision while maintaining high recall, that is, the model has better performance.

[0062] Among them, XGBoost's ROC-AUC and AP values ​​were 0.863 and 0.854, respectively, both of which were better than other models. Based on the comprehensive evaluation results, the XGBoost model performed best in all indicators and was used as the comprehensive best lightning diagnosis model in this embodiment.

[0063] Step 7: Real-world weather case verification: The generalization ability of the comprehensive optimal lightning diagnostic model is verified using data from the Global Lightning Location Network. Figure 4 This figure shows the validation results of the optimal lightning diagnostic model's application in different real-world weather events, illustrating the model's performance under various scenarios. Figure 4 a presents the lightning prediction results for a single convective cell. Figure 4 b and Figure 4 c demonstrates the prediction performance in a multi-convection single-cell environment. Figure 4 d refers to a larger area in the South China Sea (approximately 150 × 10). 4 km 2 Evaluation tests were conducted. The shaded distribution represents the brightness temperature of the BT13 infrared channel. The white squares represent lightning observed by two global lightning location networks, ENTLN (Earth Networks Total Lightning Network) and WWLLN (World Wide Lightning Location Network). (The detection efficiency of these two systems in the XX ocean area is relatively low, and the data mainly consists of high-energy ground lightning. It is reasonable that the diagnostic data covers these observations and has a larger observation range than this data.) The black circles represent lightning predicted by XGBoost.

[0064] The validation results show that the model can effectively predict lightning activity areas, has good spatial generalization ability, can maintain stable prediction performance under similar geographical conditions, and effectively predicts the lightning activity areas.

[0065] This embodiment does not require a dense network of ground-based stations at sea. It relies solely on the brightness temperature of geostationary meteorological satellites to provide kilometer-level resolution lightning risk monitoring for the deep-sea areas of the South China Sea. It is applicable to operation management and scheduling decisions in various scenarios such as offshore wind power, shipping channels, offshore engineering platforms, and fishing grounds.

[0066] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for diagnosing total flash activity areas in regions with missing lightning detection data, characterized in that, Includes the following steps: S1. Collect lightning activity data and perform spatiotemporal consistency matching; the lightning activity data includes multi-dimensional infrared channel brightness temperature data and lightning location data; S2. Identify convective cloud regions based on the multidimensional infrared channel brightness temperature data, perform lightning identification on the lightning location data, and construct a lightning tag set in the convective cloud regions based on the lightning identification results; the lightning tag set includes a positive sample set and a negative sample set; S3. Perform feature calculations based on the multidimensional infrared channel brightness temperature data to obtain multidimensional brightness temperature features, and perform feature filtering on the multidimensional brightness temperature feature set based on the lightning tag to obtain multidimensional brightness temperature representative features. S4. Based on the lightning tag set and the corresponding multi-dimensional brightness temperature representative features, perform multi-model integrated training and optimization, evaluate the model and select the comprehensive optimal lightning diagnosis model, and input the multi-dimensional infrared channel brightness temperature data into the comprehensive optimal lightning diagnosis model to obtain the lightning diagnosis result. The method for obtaining multidimensional brightness temperature features includes: Multidimensional infrared channel brightness temperature data is used as the original brightness temperature feature; Brightness temperature difference characteristics are obtained by calculating the brightness temperature difference between different infrared channels. The horizontal gradient features are obtained by calculating the horizontal spatial gradient of the brightness temperature data and brightness temperature difference of each channel. The horizontal gradient features are based on satellite grids, and the absolute value of the average spatial change rate of the original brightness temperature features or brightness temperature difference features between the satellite grid and its four neighboring satellite grids is taken as the horizontal gradient feature of the corresponding satellite grid. The number of horizontal gradient features is equal to the sum of the number of original brightness temperature features and the number of brightness temperature difference features. The multidimensional brightness temperature feature consists of the original brightness temperature feature, the brightness temperature difference feature, and the horizontal gradient brightness temperature feature.

2. The method for diagnosing total flash activity areas in regions with missing lightning detection data as described in claim 1, characterized in that, The method for constructing the lightning tag set includes: Multidimensional infrared brightness temperature data were collected by a geostationary meteorological satellite, and lightning location data were obtained by a regional three-dimensional full-flash detection system; the wavelength range of each channel of the multidimensional infrared brightness temperature data is 3.9μm~13.3μm. Cloud regions with brightness temperature values ​​less than the brightness temperature threshold within a fixed wavelength channel are defined as convective cloud regions. The minimum convex hull of lightning is calculated using a lightning radiation source attribution algorithm. Within the convective cloud region, satellite grids covered or traversed by the convex hull are marked as areas with lightning activity, while the remaining satellite grids are marked as areas without lightning activity. Satellite grids with lightning activity are used as positive samples, and satellite grids without lightning activity are used as negative samples to form a lightning label set. The minimum convex hull of lightning is used to characterize the range of lightning activity.

3. The method for diagnosing total flash activity areas in regions with missing lightning detection data as described in claim 1, characterized in that, The method for obtaining multidimensional brightness temperature representative features includes: The correlation between multidimensional brightness temperature features and lightning tag sets is calculated to obtain the lightning correlation degree of each multidimensional brightness temperature feature. The mutual redundancy among multidimensional brightness temperature features is calculated to obtain the feature redundancy degree of each multidimensional brightness temperature feature. The feature importance of each multidimensional brightness temperature feature is determined based on the lightning correlation degree and the lightning redundancy degree. The multidimensional brightness temperature features are sorted and selected according to the feature importance to obtain representative multidimensional brightness temperature features. The feature importance is positively correlated with the lightning correlation degree and negatively correlated with the feature redundancy degree.

4. The method for diagnosing total flash activity areas in regions with missing lightning detection data as described in claim 1, characterized in that, The method for multi-model ensemble training and optimization includes: Multidimensional brightness temperature representative features and corresponding lightning labels are used as a set of lightning analysis data. Lightning analysis data at all times are collected to form a lightning comprehensive set. The lightning comprehensive set is randomly divided into a training set and a test set in an 8:2 ratio. The training set is used to train the lightning activity diagnostic base model. The base models for lightning activity diagnostics include lightweight gradient boosting machines, gradient boosting decision trees, random forests, logistic regression, extreme gradient boosting, and artificial neural networks; Based on experience and theoretical analysis, the reasonable search range of key hyperparameters for each lightning activity diagnostic base model is analyzed. A multidimensional parameter grid space is constructed and an exhaustive search is performed. Combined with 5-fold cross-validation, the hyperparameters of the lightning activity diagnostic base model are systematically optimized.

5. The method for diagnosing total flash activity areas in regions with missing lightning detection data as described in claim 1, characterized in that, The method for selecting the comprehensive optimal lightning diagnostic model includes: A multi-dimensional evaluation index system is constructed. The performance of the lightning activity diagnostic base model is evaluated using training and test sets to obtain multi-dimensional evaluation indexes. Based on the multi-dimensional evaluation indexes, a comprehensive evaluation is conducted to determine the optimal lightning diagnostic model. The multi-dimensional evaluation indexes include accuracy, precision, recall, and F1 score. The generalization ability of the comprehensive optimal lightning diagnostic model was verified using actual weather events; The multidimensional infrared channel brightness temperature data is input into the comprehensive optimal lightning diagnosis model to obtain the lightning diagnosis results.

Images