Airport severe convective cloud evolution monitoring method and system based on multispectral characteristics of stationary satellite

Abstract

The invention provides an airport severe convective cloud evolution monitoring method and system based on stationary satellite multispectral characteristics, and relates to the technical field of civil aviation meteorological monitoring and early warning, and the method comprises the steps: obtaining multispectral remote sensing data of an airport monitored by a stationary satellite in real time and in a preset range; performing radiometric calibration and geometric calibration on the multispectral remote sensing data to obtain a calibrated multispectral image; extracting visible light reflectivity and infrared brightness temperature based on the calibrated multispectral image, and generating gridding feature data through spatial resampling; and based on the gridding feature data, performing semantic segmentation by adopting a full convolutional neural network integrating spatial pyramid pooling and dense up-sampling convolution so as to identify and locate a target cloud cluster with strong convection features. According to the invention, the severe convective cloud cluster is identified based on the stationary satellite data, after topographic correction and evolution prediction, graded early warning is generated in real time by fusing the airport operation state, a full-chain monitoring early warning technology is formed, and the meteorological safety guarantee capability of the airport is improved.

Description

Technical Field

[0001] This invention relates to the field of civil aviation meteorological monitoring and early warning technology, and in particular to a method and system for monitoring the evolution of strong convective cloud clusters at airports based on the multispectral characteristics of geostationary satellites. Background Technology

[0002] Severe convective weather is one of the main disaster-causing factors faced by civil aviation operations. Currently, monitoring mainly relies on weather radar, but it is limited by terrain obstruction and detection range, and its monitoring capability for nascent severe convective weather and long-distance air route areas is insufficient.

[0003] Geostationary meteorological satellites (such as the FY-4 series) provide an important data source for identifying nascent strong convection due to their high spatiotemporal resolution and multispectral observation capabilities. However, existing mainstream methods based on satellite data (such as the static threshold method and the channel brightness-temperature difference method) have significant shortcomings in data processing, making it difficult to meet the needs of refined airport early warning. Specifically, traditional methods rely on pixel-level single threshold judgments and cannot integrate multispectral and multi-scale contextual information, which may lead to high false alarm rates (such as misclassifying high-altitude cirrus clouds as convective clouds) and false negative rates, and insufficient accuracy in locating cloud boundaries and outlines. Furthermore, these methods are mostly limited to identification at the current moment and lack the ability to use continuous time-series data for intelligent tracking and evolution modeling, which may prevent quantitative prediction of future intensity changes and movement of cloud clusters. The forecasts suffer from several problems: insufficient lead time for early warnings; a lack of deep integration of high-precision terrain data in data processing, failing to quantify terrain undulations, slope aspects, and relative distances to airports into calculable parameters; inability to correct for spatial differences in terrain effects in cloud evolution, potentially leading to large deviations in forecasts in complex terrain areas; and reliance on forecasters' subjective judgment in generating early warning information, lacking models that automatically and quantitatively integrate corrected meteorological forecasts with real-time airport operational status (such as flight schedules, runway occupancy, and airspace traffic), potentially compromising the relevance, timeliness, and consistency of early warning instructions. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a method and system for monitoring the evolution of strong convective clouds at airports based on the multispectral characteristics of geostationary satellites. The method identifies strong convective clouds based on geostationary satellite data, and after terrain correction and evolution prediction, integrates the airport's operational status to generate graded early warnings in real time, forming a full-chain monitoring and early warning technology, which improves the airport's meteorological safety assurance capabilities.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a method for monitoring the evolution of strong convective cloud clusters at airports based on geostationary satellite multispectral characteristics, the method comprising: Acquire real-time multispectral remote sensing data of airports and preset areas monitored by geostationary satellites; Radiometric and geometric calibration processing is performed on multispectral remote sensing data to obtain calibrated multispectral images; visible light reflectance and infrared brightness temperature are extracted based on the calibrated multispectral images, and gridded feature data is generated through spatial resampling; based on the gridded feature data, a fully convolutional neural network integrating spatial pyramid pooling and dense upsampling convolution is used to perform semantic segmentation to identify and locate target cloud clusters with strong convection characteristics. For the target cloud cluster, a polygonal monitoring area is constructed based on the airport's influence zone and subjected to multi-grid processing to obtain terrain-adaptive grid weight parameters. Based on the terrain-adaptive grid weight parameters, an ensemble learning sliding step prediction method is used to obtain normalized spectral feature time-series data of the target cloud cluster. Based on the normalized spectral feature time-series data, a time-series evolution model is constructed using the sliding window method to generate initial state projection values ​​for the future state of the cloud cluster. The initial state projection values ​​are spatially corrected using the terrain-adaptive grid weight parameters to obtain the corrected future evolution trend of the cloud cluster and its potential impact on the airport. Based on the corrected future evolution trend of the cloud cluster and its potential impact on the airport, combined with the airport's real-time operational status parameters, a graded early warning instruction is generated, which includes the impact level, the expected period of impact, and recommended handling measures, and is then released in real time.

[0006] Secondly, an airport strong convective cloud evolution monitoring system based on geostationary satellite multispectral characteristics includes: The acquisition module is used to acquire real-time multispectral remote sensing data of airports and preset ranges monitored by geostationary satellites. The intelligent recognition module is used to perform radiometric and geometric calibration on multispectral remote sensing data to obtain calibrated multispectral images. Based on the calibrated multispectral images, visible light reflectance and infrared brightness temperature are extracted and generated into gridded feature data through spatial resampling. Based on the gridded feature data, a fully convolutional neural network integrating spatial pyramid pooling and dense upsampling convolution is used to perform semantic segmentation to identify and locate target cloud clusters with strong convection characteristics. The correction module is used to construct a polygonal monitoring area based on the airport's influence zone for the target cloud cluster and perform multi-grid processing to obtain terrain-adaptive grid weight parameters; based on the terrain-adaptive grid weight parameters, an ensemble learning sliding step prediction method is used to obtain normalized spectral feature time-series data of the target cloud cluster; based on the normalized spectral feature time-series data, a time-series evolution model is constructed using the sliding window method to generate initial state projection values ​​for the future state of the cloud cluster; and the initial state projection values ​​are spatially corrected by combining the terrain-adaptive grid weight parameters to obtain the corrected future evolution trend of the cloud cluster and its potential impact on the airport. The early warning module is used to generate graded early warning instructions, including the impact level, the expected period of impact, and recommended handling measures, based on the corrected future evolution trend of the cloud cluster and its potential impact on the airport, combined with the airport's real-time operational status parameters, and to issue them in real time.

[0007] Thirdly, a computing device, comprising: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0008] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0009] The above-described solution of the present invention has at least the following beneficial effects: This process acquires real-time multispectral remote sensing data of the airport and a predetermined area from geostationary satellites, providing a continuous multispectral data source covering the target area based on the characteristics of geostationary satellite data. It ensures the real-time nature and dimensional integrity of the initial data, providing a suitable foundation for subsequent data processing. Multispectral remote sensing data is processed through radiometric and geometric calibration to eliminate systematic errors and spatial location biases, ensuring the data quality of the calibrated images. Visible reflectance and infrared brightness temperature are extracted, and spatial resampling generates gridded feature data, transforming unstructured images into structured data and improving the data's compatibility with subsequent models. A fully convolutional neural network integrating spatial pyramid pooling and dense upsampling convolutions is used for semantic segmentation, achieving comprehensive extraction of multi-scale contextual features and restoration of spatial details, optimizing the data analysis process for cloud cluster identification, and improving the targeting of cloud cluster identification and localization. A polygonal monitoring area is constructed based on the airport's influence zone and multi-griddged, defining the monitoring range while achieving hierarchical data structuring, providing a suitable carrier for terrain-related data processing. Terrain-adaptive grid weight parameters are generated, transforming terrain information into a usable data structure. The system calculates quantitative data to adapt data processing to the actual geographical environment; it employs an ensemble learning sliding step prediction method to obtain normalized spectral feature time-series data, integrating the advantages of multiple models and enriching the dimensions of time-series data through sliding step sampling, thereby improving the reliability and completeness of the time-series data; it utilizes the sliding window method to construct a time-series evolution model, achieving standardized sorting and orderly analysis of time-series data, supporting the data processing flow for cloud cluster future state projection; it combines terrain-adaptive grid weight parameters for spatial correction, achieving deep integration of time-series projection data and terrain spatial differences, improving the adaptability of evolution data to the actual spatial environment; it integrates the corrected cloud cluster information with airport real-time operational status parameters, achieving cross-dimensional integration of meteorological and operational data, making early warning-related data processing more aligned with actual airport operations; it generates hierarchical early warning instructions containing multiple elements through systematic data processing, achieving standardized and structured output of early warning information, ensuring the pertinence and consistency of early warning instructions; it issues hierarchical early warning instructions in real time, achieving efficient flow of processed early warning information, ensuring timely access and application by relevant departments, and ensuring the orderly conduct of airport operational response work. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the method for monitoring the evolution of strong convective clouds at airports based on geostationary satellite multispectral features, provided by an embodiment of the present invention.

[0011] Figure 2 This is a schematic diagram of an airport strong convective cloud evolution monitoring system based on geostationary satellite multispectral features, provided by an embodiment of the present invention. Detailed Implementation

[0012] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0013] like Figure 1 As shown, embodiments of the present invention propose a method for monitoring the evolution of strong convective clouds at airports based on geostationary satellite multispectral characteristics. The method includes the following steps: Step 100: Obtain multispectral remote sensing data of the airport and the preset range monitored in real time by geostationary satellites; Step 200: Radiometric and geometric calibration are performed on the multispectral remote sensing data to obtain calibrated multispectral images; visible light reflectance and infrared brightness temperature are extracted based on the calibrated multispectral images, and gridded feature data is generated through spatial resampling; based on the gridded feature data, a fully convolutional neural network integrating spatial pyramid pooling and dense upsampling convolution is used to perform semantic segmentation to identify and locate target cloud clusters with strong convection characteristics. Step 300: For the target cloud cluster, a polygonal monitoring area is constructed based on the airport's influence zone and subjected to multi-grid processing to obtain terrain-adaptive grid weight parameters; based on the terrain-adaptive grid weight parameters, an ensemble learning sliding step prediction method is used to obtain normalized spectral feature time-series data of the target cloud cluster; based on the normalized spectral feature time-series data, a time-series evolution model is constructed using the sliding window method to generate initial state projection values ​​for the future state of the cloud cluster; combined with the terrain-adaptive grid weight parameters, the initial state projection values ​​are spatially corrected to obtain the corrected future evolution trend of the cloud cluster and its potential impact on the airport; Step 400: Based on the corrected future evolution trend of the cloud cluster and its potential impact on the airport, combined with the airport's real-time operational status parameters, a graded early warning instruction is obtained, which includes the impact level, the expected period of impact, and recommended handling measures, and is then released in real time.

[0014] In this embodiment of the invention, multispectral remote sensing data of the airport and a preset range are acquired in real time, ensuring the timeliness and relevance of the data source and providing a basic data source for subsequent processing. The preset range can focus on the core monitoring area, reduce interference from irrelevant data, and improve processing efficiency. Radiometric calibration and geometric calibration ensure the accuracy of multispectral images, providing a reliable foundation for feature extraction. Spatial resampling generates gridded feature data, making the feature distribution more regular and facilitating subsequent network processing. A fully convolutional neural network integrating spatial pyramid pooling and dense upsampling convolution can fully mine spectral and spatial features, improve the effectiveness of target cloud cluster identification and positioning, and achieve feature extraction and effective cloud cluster identification and positioning. Multiple grids The processing, combined with terrain-adaptive grid weight parameters, ensures that the processing conforms to actual terrain conditions and improves spatial adaptability. The ensemble learning sliding step prediction method can stably acquire normalized spectral feature time-series data, ensuring the reliability of the time-series data. The time-series evolution model constructed by the sliding window method can effectively capture the evolution patterns of cloud clusters, and spatial correction can optimize the initial inference results, making the future evolution trend and potential impact information of cloud clusters more consistent with the actual spatial distribution. The integration of real-time airport operation status parameters makes the early warning instructions more adaptable to the actual operational needs of the airport, and the hierarchical setting makes the early warning more layered. The timely release of support early warning information facilitates timely response and handling by relevant departments, improving the pertinence and timeliness of airport operation support.

[0015] In a preferred embodiment of the present invention, step 100, which involves acquiring multispectral remote sensing data of the airport and a preset range monitored in real time by a geostationary satellite, includes: Step 101: Obtain the full-disk first-level data product from the multi-channel scanning imager of a geostationary meteorological satellite. Specifically, this includes: First, selecting a geostationary meteorological satellite with high spatiotemporal resolution and multispectral observation capabilities, preferably the FY-4 series geostationary meteorological satellite. The multi-channel scanning imager on this satellite can achieve continuous observation of the entire disk area. Then, through the satellite data receiving system, the full-disk first-level data product output by the multi-channel scanning imager is acquired in real time. This level of data product has not undergone complex preprocessing and completely preserves the original radiation information and spatial location information during the satellite observation process. It can provide basic support for the subsequent extraction of spectral data of the target area. At the same time, its full-disk coverage characteristics can effectively expand the monitoring range, covering the airport and the surrounding pre-set long-distance flight path area, providing data support for the early identification of nascent strong convection.

[0016] Step 102 involves extracting the raw observation counts of the infrared split-window channel and water vapor channel covering the airport and a preset range from the full-disk primary data product. Specifically, this includes: First, defining a preset monitoring range centered on the airport based on its actual geographical coordinates. This range must cover the airport runway, terminal building, approach routes, and surrounding areas prone to strong convective clouds. Then, based on the geographical boundaries of this preset monitoring range, cropping the corresponding dataset from the full-disk primary data product. Further, combining the spectral response characteristics of strong convective clouds, the raw observation counts of the infrared split-window channel and water vapor channel within this region are extracted. The infrared split-window channel effectively reflects the temperature characteristics of the cloud, while the water vapor channel captures the water vapor content distribution of the cloud. Both are core spectral channels for identifying strong convective clouds. By focusing on the extraction of data from these two channels, redundant interference from irrelevant channel data can be reduced, while retaining the key raw observation information required for identifying strong convective clouds, laying a data foundation for subsequent radiometric calibration and feature analysis.

[0017] Step 103: Using the radiometric calibration parameters corresponding to each channel, the original observation counts are converted into brightness and temperature data to obtain radiometrically calibrated channel data. Specifically, this includes: to address the insufficient recognition accuracy caused by inconsistent data radiometric characteristics, and to correct the impact of geometric deviations in the radiation transmission path caused by Earth's curvature on the calibration results, a radiometric calibration process combined with an Earth curvature radiation path geometric compensation algorithm is used to achieve accurate quantization and conversion of the original observation data, giving the data physical meaning. Specifically, firstly, the radiometric calibration parameters corresponding one-to-one with the infrared split-window channel and the water vapor channel are retrieved. These parameters are determined by ground calibration and on-orbit calibration results before satellite launch and include the gain coefficient and offset of each channel, which can initially eliminate systematic errors caused by differences in the satellite observation instrument's own response and changes in the observation environment. Further, an Earth curvature radiation path geometric compensation algorithm is introduced for deviation correction. First, the geographical coordinates of the current observation area and the satellite observation zenith angle are obtained. Then, by querying a preset geophysical parameter database, the Earth curvature radius parameter corresponding to the observation area is obtained, combined with the satellite... The geometric deviation of the radiation transmission path is calculated from the satellite orbital altitude data. Based on this geometric deviation, the previously retrieved radiometric calibration parameters are compensated and corrected. During the correction process, the curvature coefficient of the radiation path is first determined based on the ratio of the Earth's radius of curvature to the satellite's orbital altitude. Then, the gain coefficient and offset are dynamically adjusted pixel by pixel using the curvature coefficient to complete the geometric compensation optimization of the calibration parameters. Subsequently, the original observation counts are substituted into the compensated and corrected radiometric calibration models of each channel to complete the conversion of the original observation counts into brightness and temperature data. Specifically, the original counts of the infrared split window channel are converted into the corresponding cloud brightness and temperature, and the original counts of the water vapor channel are converted into brightness and temperature data reflecting water vapor distribution. Through the above combined radiometric calibration and geometric compensation processing, radiometrically calibrated channel data is obtained. This data not only has a unified radiometric quantification standard but also eliminates the calibration error caused by the Earth's curvature radiation path deviation, ensuring the comparability of data between different time periods and different channels. This provides a more reliable quantitative basis for subsequent multispectral information fusion analysis and strong convective cloud feature extraction.

[0018] Step 104: Based on the standard Earth coordinate system, perform geometric fine correction and resampling on the radiometrically calibrated channel data to obtain equal latitude and longitude projection grid data with uniform spatial resolution, which serves as multispectral remote sensing data. Specifically, this includes: First, determining the standard Earth coordinate system as the reference coordinate system for geometric correction, ensuring that the spatial location of the data matches the actual geographic coordinates; then, based on this standard Earth coordinate system, using ground control point data, performing geometric fine correction processing on the radiometrically calibrated channel data to correct geometric distortions caused by factors such as satellite attitude changes, orbital offsets, and Earth curvature during satellite observation. To ensure that the spatial location of each pixel in the channel data accurately corresponds to the actual geographical area, a unified spatial resolution is set according to the refined monitoring requirements of strong convective cloud clusters. The channel data after geometric correction is then resampled to ensure that the data from different channels have a consistent spatial resolution. Finally, isotropic projection grid data is generated as multispectral remote sensing data required for subsequent processing. This grid data not only ensures the accuracy of spatial location and the spatial consistency of data from each channel, but also facilitates subsequent grid-level spatial analysis in conjunction with high-precision terrain data, providing a foundation for correcting the terrain effects of cloud cluster evolution.

[0019] In this embodiment of the invention, the full-disk first-level data product of the geostationary meteorological satellite multi-channel scanning imager is acquired, ensuring the integrity and originality of the data source and providing a comprehensive basic data source for subsequent data extraction from specific areas. The multi-channel characteristics can cover the multi-spectral information required for strong convection monitoring, improving the applicability of the data. The original observation counts of the infrared split-window channel and the water vapor channel are extracted in a targeted manner, focusing on the core spectral channels for monitoring strong convective cloud clusters, reducing redundant interference from irrelevant channel data, and improving the targeting and efficiency of data processing. Directly acquiring the original observation counts can retain the most original observation information of the data, providing a reliable foundation for subsequent calibration processing. The process involves converting raw observation counts into brightness and temperature data using corresponding channel radiometric calibration parameters, eliminating systematic errors caused by differences in instrument response and the observation environment, and giving the data physical meaning. It also ensures the radiometric consistency of data from each channel, providing a quantitative basis for subsequent cross-channel data fusion and analysis. Geometric calibration based on the standard Earth coordinate system corrects geometric distortions and ensures accurate matching between the data's spatial location and actual geographic coordinates. Resampling generates isotropic latitude and longitude projection grid data and unifies spatial resolution, providing a unified spatial benchmark for subsequent cross-time period data comparison and spatial analysis, thus improving the spatial consistency and usability of the data.

[0020] In a preferred embodiment of the present invention, step 200 involves radiometric and geometric calibration of the multispectral remote sensing data to obtain a calibrated multispectral image; extracting visible light reflectance and infrared brightness temperature from the calibrated multispectral image, and generating gridded feature data through spatial resampling; and using the gridded feature data, performing semantic segmentation to identify and locate target cloud clusters with strong convective characteristics using a fully convolutional neural network integrating spatial pyramid pooling and dense upsampling convolution, including: Step 201: Based on the multispectral remote sensing data, reflectance calibration is performed on the visible light bands included, and brightness-temperature conversion is performed on the infrared split-window channel and water vapor channel to obtain radiometrically calibrated multispectral data. Specifically, this includes: First, determining the band types included in the data based on the multispectral remote sensing data, distinguishing the visible light band, infrared split-window channel, and water vapor channel; Targeted calibration processing is carried out for the observation characteristics of different bands. For the visible light band, a preset reflectance calibration process is used, combined with the calibration coefficients of the satellite payload to complete the conversion of the original observation data to reflectance, so as to characterize the cloud's reflection characteristics of visible light; For the infrared split-window channel and water vapor channel, the radiometric calibration logic is continued, and the conversion of the original data to brightness-temperature is completed through the corresponding calibration parameters to accurately reflect the thermal radiation characteristics of the cloud; Through the above differentiated radiometric calibration operations, radiometrically calibrated multispectral data is finally obtained. This data has a unified radiometric quantification standard, which can ensure the comparability of data from different bands and lay a solid foundation for subsequent multispectral information fusion analysis.

[0021] Step 202: Based on the radiometrically calibrated multispectral data, geometric fine correction and reprojection are performed according to a standard geographic reference system to obtain a calibrated multispectral image. Specifically, this includes: First, using the radiometrically calibrated multispectral data as the processing object, a standard geographic reference system is selected as the benchmark for geometric correction. This standard geographic reference system includes a unified coordinate system and projection method, ensuring accurate matching between the image's spatial location and the actual geographic area. Then, using a preset ground control point dataset, geometric fine correction is performed on the radiometrically calibrated multispectral data to correct geometric distortions caused by satellite attitude changes, orbital offsets, and Earth curvature during satellite observation. Based on this, according to the geographical features of the airport and the preset monitoring range, the calibrated multispectral data is reprojected to adapt the image's projection method to the needs of subsequent spatial analysis. Finally, a calibrated multispectral image is obtained. This image not only retains spectral information but also possesses reliable spatial location attributes, providing an accurate spatial carrier for subsequent feature extraction and cloud cluster positioning.

[0022] Step 203 involves extracting visible light reflectance and infrared brightness temperature from the calibrated multispectral image and performing spatial interpolation resampling to generate gridded feature data with uniform spatial resolution. Specifically, this includes: First, using the calibrated multispectral image as the data source, extracting core feature parameters closely related to the identification of strong convective clouds, namely visible light reflectance and infrared brightness temperature. These two parameters characterize cloud attributes from the perspectives of reflectance and thermal radiation, respectively, and are key criteria for distinguishing between strong convective clouds and ordinary clouds. Then, considering the potential spatial resolution differences between different spectral bands, spatial interpolation resampling is performed on the extracted visible light reflectance and infrared brightness temperature data. A preset spatial resolution is selected as a unified standard, and interpolation algorithms are used to fill data gaps, ensuring consistent spatial resolution for both types of feature data. Finally, gridded feature data with uniform spatial resolution is generated. This data organizes discrete feature parameters into a structured grid, preserving accurate spectral feature information and improving the data's adaptability to subsequent neural network models, providing data support for efficient cloud identification.

[0023] Step 204: Construct a fully convolutional neural network semantic segmentation model. The encoder of this model integrates a spatial pyramid pooling structure to extract multi-scale contextual features, while the decoder employs a dense upsampling convolutional structure to recover spatial details. Specifically, this includes: first, model construction, establishing the basic architecture of the fully convolutional neural network; designing the network hierarchy based on the requirements for extracting features from strong convective cloud clusters; determining the number and connection methods of convolutional and pooling layers; and defining the core functional division between the encoder and decoder. The encoder is responsible for layer-by-layer feature extraction and dimensionality compression of the input gridded feature data, while the decoder is responsible for layer-by-layer upsampling and spatial restoration of the high-dimensional features output by the encoder. The encoder integrates a spatial pyramid pooling structure, specifically designing pooling windows of various scales based on the common scale range of strong convective cloud clusters. These pooling operations at different scales process the feature maps output by the encoder in parallel, capturing contextual features at different spatial scales. This enables the identification of both the overall features of large-scale strong convective clouds and the capture of local details at small scales, achieving comprehensive extraction of multi-scale features. The decoder employs a dense upsampling convolutional structure. This structure tightly connects adjacent upsampling layers with convolutional layers, allowing the output features of each upsampled layer to be fused with the features of multiple preceding convolutional layers. This effectively recovers the spatial detail information compressed by the encoder, improves the accuracy of cloud boundary and contour reconstruction, and compensates for the lack of detail loss in traditional semantic segmentation models. Through this architectural design, a fully convolutional neural network semantic segmentation model is finally constructed. This model possesses multi-scale feature extraction and accurate detail reconstruction capabilities, laying the foundation for subsequent model training and efficient and accurate semantic segmentation of strong convective clouds.

[0024] Step 205 involves training the fully convolutional neural network semantic segmentation model using sample data containing annotations of strong convective cloud clusters. The gridded feature data is input into the trained model to obtain the category probability of each pixel. Specifically, this includes: First, preparing a training dataset containing a large number of multispectral image samples covering different meteorological conditions and time periods, with each sample annotated with the actual area of ​​a strong convective cloud cluster, ensuring the diversity and accuracy of the training samples. The training dataset is then input into the constructed fully convolutional neural network semantic segmentation model, and the model parameters are iteratively updated using a preset training strategy. This allows the model to gradually learn the differences between strong convective clouds and ordinary clouds in features such as visible light reflectance and infrared brightness temperature, improving the model's ability to identify strong convective clouds. After the model training is complete and meets the preset convergence conditions, the gridded feature data is input into the trained model. The model uses forward inference to determine the category attribute of each pixel and outputs the category probability of each pixel belonging to a strong convective cloud cluster. This category probability provides a quantitative basis for subsequent cloud cluster selection, effectively improving the reliability of cloud cluster identification.

[0025] Step 206: Based on the category probability of each pixel, connected component analysis and threshold discrimination methods are used to identify and extract target cloud clusters with strong convective characteristics within a preset range, and determine their spatial location and contour information. Specifically, this includes: First, threshold discrimination and preliminary screening of potential regions are performed. Taking the category probability of each pixel as the processing object, combined with the statistical results of the category probability of a large number of historical samples of strong convective clouds, and referring to the actual needs of civil aviation meteorological monitoring, a reasonable probability threshold is set. This threshold is usually in the range of 0.6 to 0.8, and the final threshold is determined through multiple sample verifications and optimizations. The threshold discrimination method is used to determine the category probability of all pixels one by one. Pixels with a category probability higher than the set threshold are marked as potential strong convective cloud cluster pixels, and pixels with a category probability lower than or equal to the threshold are marked as background pixels. In this way, a preliminary strong convective cloud cluster pixel region composed of continuous potential pixels is selected, laying the foundation for subsequent identification.

[0026] Based on this, connected component analysis and noise removal are carried out. Connected component analysis is used to process the initially screened pixel regions. First, the criteria for determining adjacent pixels are set. The 4-neighborhood connectivity rule is used, that is, pixels adjacent in the four directions (up, down, left, and right) are considered as valid neighbors. Alternatively, the 8-neighborhood connectivity rule is selected based on the spatial distribution characteristics of the cloud clusters, that is, pixels adjacent in the diagonal direction are included. By scanning the entire initially screened area, pixels that are adjacent and whose category attribute is potential strong convective cloud clusters are aggregated into complete connected regions. At the same time, the number of pixels contained in each connected region is counted. A threshold for determining noise pixel regions is set. Connected regions containing fewer than the preset threshold are determined as isolated noise pixel regions. The preset threshold is determined based on the minimum pixel size of historical cloud cluster samples, which is usually 3 to 5 pixels. Regions determined to be noise are removed to ensure that the identified cloud cluster regions have structural integrity and continuity.

[0027] Subsequently, geographic range matching and target cloud cluster screening are performed. Geographic boundary data of the airport and the preset monitoring range are retrieved in advance. This data is a closed area defined by latitude and longitude coordinates, including the airport's core operating area and the surrounding preset influence extension area. Each complete connected area retained after connected domain analysis is spatially matched with the preset geographic boundary data. By calculating the latitude and longitude coordinates of the center point of each connected area, it is determined whether the center point falls within the preset geographic boundary range, or the percentage of overlap between the connected area and the preset geographic boundary range is calculated. When the percentage of overlap is greater than the preset ratio, it is considered to meet the range requirements. From all connected areas, areas that meet the geographic range requirements are screened out and identified as target cloud clusters with strong convective characteristics, ensuring that the identification results focus on cloud cluster areas that may affect the airport.

[0028] Simultaneously, spatial location and contour information extraction is performed. A contour tracking algorithm is used to extract the boundaries of the connected regions of the identified target cloud cluster. The edge pixels of the connected regions are tracked point by point, and the latitude and longitude coordinates of each edge pixel are recorded to form a continuous sequence of boundary coordinates. The contour information of the target cloud cluster is determined by the boundary coordinate sequence. This contour information is presented in the form of an ordered set of coordinate points, which can accurately represent the edge morphology of the cloud cluster. At the same time, the latitude and longitude coordinates of the center point of the connected regions of the target cloud cluster are calculated, or the circumscribed rectangle of the cloud cluster is generated by fitting the boundary coordinates. The spatial location of the cloud cluster is represented by the combination of the center point coordinates and the size of the circumscribed rectangle. Finally, the identification and localization of the target cloud cluster are completed, and the spatial location parameters and contour coordinate information of the target cloud cluster are output, providing spatial basis data for subsequent cloud cluster evolution tracking and impact assessment.

[0029] In this embodiment of the invention, differentiated radiometric calibration is performed on different bands of multispectral remote sensing data. Reflectance calibration is performed on the visible light band, and brightness-temperature conversion is performed on the infrared split-window channel and water vapor channel to ensure accurate calibration of the radiometric characteristics of different types of spectral data. This ensures that the radiometrically calibrated multispectral data has a unified physical meaning, laying a reliable foundation for subsequent multi-band data fusion analysis. Geometric fine correction and reprojection are performed based on a standard geographic reference system to effectively correct geometric distortions in the multispectral data, ensuring that the spatial location of the calibrated multispectral image matches the actual geographic area. Reprojection unifies the spatial reference of the image, improving its spatial consistency and facilitating subsequent cross-time period and cross-data source comparative analysis. Visible reflectance and infrared brightness-temperature core features closely related to the identification of strong convective clouds are extracted from the calibrated image. Spatial interpolation and resampling generate gridded feature data with unified spatial resolution, making the feature data distribution more regular and reducing spatial resolution. Interference caused by differences; the fully convolutional neural network encoder integrates a spatial pyramid pooling structure, which can fully capture contextual features at different scales and improve the ability to extract features of strong convective clouds of different sizes; the decoder adopts a dense upsampling convolutional structure, which can effectively restore spatial detail information, enhance the accuracy of cloud boundary and contour restoration, and optimize the basic model architecture of semantic segmentation; the model is trained using sample data with strong convective cloud annotations, so that the model can fully learn the feature patterns of strong convective clouds and improve the model's recognition adaptability to target clouds; gridded feature data is input into the trained model to obtain pixel-level category probabilities, providing a quantitative basis for subsequent cloud identification; connected component analysis can effectively aggregate similar pixels and filter out complete strong convective cloud regions; threshold discrimination is combined to filter target clouds and reduce interference from non-strong convective cloud pixels; at the same time, the spatial location and contour information of target clouds are determined, providing spatial basis data for subsequent cloud evolution monitoring and impact assessment.

[0030] In a preferred embodiment of the present invention, step 300 involves, for the target cloud cluster, constructing a polygonal monitoring area based on the airport impact zone and performing multi-grid processing to obtain terrain-adaptive grid weight parameters; based on the terrain-adaptive grid weight parameters, using an ensemble learning sliding step size prediction method to obtain normalized spectral feature time-series data of the target cloud cluster; based on the normalized spectral feature time-series data, using a sliding window method to construct a time-series evolution model to generate initial state projection values ​​for the future state of the cloud cluster; and combining the terrain-adaptive grid weight parameters to spatially correct the initial state projection values ​​to obtain the corrected future evolution trend of the cloud cluster and its potential impact on the airport, including: Step 301: For the target cloud cluster, based on the airport's geographical location and flight path distribution, construct a polygonal monitoring area covering the potential impact area within a preset spatial buffer distance. Specifically, this includes: First, based on the spatial coordinates and outline boundary information of the target cloud cluster, combined with the geographical location data of the airport's core operating area (including key facilities such as runways, terminals, control towers, and approach navigation beacons), and simultaneously retrieving the distribution and orientation data of the airport's arrival glide slope, departure climb section, and main flight paths, through spatial overlay analysis, determining key areas such as flight takeoff and landing paths and ground operating areas that may be affected during the cloud cluster's movement; based on this, setting a reasonable preset spatial buffer distance. The setting of this buffer distance needs to comprehensively consider the typical impact radius of strong convective clouds, historical movement speed data, and the airport's advance warning requirements. It is usually set to 20 to 50 kilometers to ensure full coverage of the airport-related areas that may be affected by the cloud's movement in the next 1 to 2 hours. Subsequently, based on the determined geographical location of the core area, the distribution range of air routes, and the preset spatial buffer distance parameters, a closed polygonal monitoring area is constructed using a convex hull algorithm or a polygon fitting algorithm. This area can accurately define the potential impact range of the target cloud on the airport, avoiding redundant data interference caused by an excessively large monitoring range, and preventing the omission of key impact areas caused by an excessively small monitoring range.

[0031] Step 302: Perform multi-mesh processing on the polygonal monitoring area. This multi-mesh processing includes generating a nested multi-level mesh structure based on at least two different spatial resolutions. Specifically, it includes: First, taking the polygonal monitoring area as the processing object, determining at least two different spatial resolution parameters. The lower spatial resolution parameter is set to 2 kilometers, corresponding to a pixel ground sampling distance of 2 kilometers × 2 kilometers. This resolution is suitable for analyzing the overall macroscopic cloud evolution of the monitoring area and can quickly capture the movement and intensity changes of large-scale cloud clusters. The higher spatial resolution parameter is set to 500 meters, corresponding to a pixel ground sampling distance of 500 meters × 500 meters. This resolution is suitable for the core operating area of ​​the airport. To meet the microscopic analysis needs of the current densely distributed areas of cloud clusters, it can capture cloud cluster boundary details and local intensity changes. Subsequently, based on the selected different spatial resolutions, a geographic coordinate system-based grid structure is generated sequentially within the polygonal monitoring area. During the generation process, it is ensured that the coordinate systems of the high-resolution grid and the low-resolution grid are consistent, so that the grids of different resolutions are nested to form a multi-level grid structure. Among them, the high-resolution grid mainly covers the core area of ​​the airport and the area where the cloud cluster is currently located, while the low-resolution grid covers the entire polygonal monitoring area. This nested structure allows subsequent processing to grasp the evolution trend of the cloud cluster from a global perspective and to carry out detailed analysis in key areas, improving the adaptability of the grid structure to subsequent terrain fusion and time series analysis.

[0032] Step 303: Obtain the spatial extent definition data of the multi-layer grid structure; based on the spatial extent definition data, extract the elevation data of the corresponding region from the digital elevation model database, specifically including: First, extract the spatial extent definition data of each multi-level grid structure. This data contains core information such as the boundary latitude and longitude coordinates of each level of grid, the row and column numbers of the grid cells, the coordinate range of each grid cell, and the spatial resolution. This information can define the specific geographical area covered by each level of grid; Based on this, based on these spatial extent definition data, extract the elevation data of the corresponding geographical area from the pre-constructed high-precision digital elevation model database. The construction and implementation process of the high-precision digital elevation model is as follows: First, collect multi-source terrain data sources of the target area, including satellite remote sensing stereo imagery, aerial photogrammetry data, and ground measured elevation point data. Preprocess the collected multi-source data, sequentially completing data correction, noise removal, and... Coordinate registration ensures spatial consistency across data sources. Subsequently, a 3D terrain reconstruction method is used to fuse the preprocessed data sources, filling data gaps with spatial interpolation algorithms to generate continuous terrain elevation raster data. The raster data is then validated and optimized for accuracy, eliminating outlier values ​​and correcting deviations to ultimately form a high-precision digital elevation model. This model is divided and stored according to geographical regions, forming a digital elevation model database with a resolution of at least 30 meters to ensure the accuracy of the terrain data. During elevation data extraction, the coordinate range of grid cells is matched with the coordinate range of the digital elevation model, ensuring that each cell at each level of the grid matches the corresponding elevation data. For single grid cells containing multiple elevation values, the arithmetic mean method is used to calculate the average elevation of the grid cell as the final matching result. This achieves the integration of the grid structure with the basic terrain data, establishing a data framework for terrain impact analysis.

[0033] Step 304: Based on the elevation data, calculate the topographic relief and aspect characteristic values ​​of each grid cell; based on the topographic relief and aspect characteristic values, and integrating the spatial distance information between each grid cell and the airport core area, calculate and generate the topographic influence weight coefficient for each grid cell. Specifically, this includes: First, based on the elevation data, perform topographic feature calculations for each grid cell. The topographic relief is obtained by calculating the difference between the maximum and minimum elevation values ​​within each grid cell and its adjacent 3×3 grids, thus reflecting the severity of topographic undulation in the area; the aspect characteristic value is obtained by calculating the direction angle of the elevation gradient within the grid cell, thus characterizing the orientation of the terrain and quantifying the core topographic attributes; on this basis, using a spatial distance calculation algorithm, obtain the topographic influence weight coefficient between each grid cell and the airport core area. The spatial distance information includes the straight-line distance from the center point of the grid cell to the center point of the airport core area, and the shortest path distance after considering terrain obstruction. These two distance information serve as the quantitative basis for spatial location association. Subsequently, this distance information is integrated with the aforementioned terrain undulation and slope aspect characteristics. Based on the importance of each parameter to cloud evolution, the weight ratio of each parameter is determined through expert experience assignment or statistical regression analysis. Specifically, the weight ratio of terrain undulation is 40%, the weight ratio of slope aspect characteristics is 30%, and the weight ratio of spatial distance information is 30%. Then, the terrain influence weight coefficient of each grid cell is generated by linear weighted summation. This coefficient can accurately characterize the terrain conditions and spatial relationship with the airport at different grid cells, and the degree of differentiated influence on cloud evolution.

[0034] Step 305 involves normalizing and fusing the terrain influence weight coefficients of all grid levels to obtain terrain-adaptive grid weight parameters. Specifically, this includes: First, normalizing the terrain influence weight coefficients of each grid level using a min-max normalization method to map the weight coefficients of different grid levels to a uniform range of 0 to 1, eliminating the incomparability of weight values ​​caused by spatial resolution differences and ensuring the comparability of weight coefficients between different grid levels. Based on this, a multi-scale data weighted fusion strategy is adopted to fuse the normalized weight coefficients of each grid level. The fusion weights are determined according to the resolution accuracy of different grid levels, with high-resolution grids accounting for 60% and low-resolution grids accounting for 40%. Multi-scale weight information is fused through weighted average calculation, fully preserving the macroscopic terrain distribution patterns and microscopic terrain details contained in the weights of different grid levels. Finally, terrain-adaptive grid weight parameters are obtained, which can adapt to the subsequent processing needs of cloud data at different scales, laying the foundation for incorporating terrain-differentiated influences.

[0035] Step 306: Based on the spatial location and contour information of the target cloud cluster, extract the infrared and visible light channel spectral values ​​within the corresponding spatial range from the continuous time-series images of the multispectral remote sensing data to obtain the original spectral feature data of the target cloud cluster. Specifically, this includes: First, based on the spatial location coordinates and contour boundary information of the target cloud cluster, determine the specific spatial range from which the spectral data needs to be extracted by spatial range clipping. This range is extended outward by one grid cell based on the contour boundary of the target cloud cluster to ensure complete coverage of the edge region of the target cloud cluster. On this basis, retrieve the continuous time-series images corresponding to the multispectral remote sensing data. The time interval of these time-series images is consistent with the satellite observation cycle. Typically, each frame is 15 minutes long, containing observational information of the target cloud cluster at multiple consecutive moments. Subsequently, using spatial region matching and image cropping algorithms based on latitude and longitude coordinates, the infrared and visible light spectral values ​​within the aforementioned spatial range are extracted frame by frame from the continuous temporal images. During the extraction process, it is ensured that the spectral values ​​of the corresponding spatial range in each frame are completely extracted without omission or redundancy. Finally, these frame-by-frame extracted spectral values ​​are arranged in time-stamp order to form an array of original spectral feature data distributed according to time series. This data completely preserves the pattern of spectral characteristics of the target cloud cluster changing over time, providing a continuous and complete data source for subsequent time-series prediction.

[0036] Step 307: Using the terrain-adaptive grid weight parameters, perform spatial weighting and normalization calculations on the original spectral feature data of the target cloud cluster to obtain normalized spectral feature data of the target cloud cluster. Specifically, this includes: First, calling the terrain-adaptive grid weight parameters and matching the weight coefficients corresponding to each grid cell using the coordinates of the grid cells to ensure that the weight parameters accurately correspond to the grid cells containing the spectral data; Based on this, perform spatial weighting calculations on the obtained original spectral feature data of the target cloud cluster and the weight coefficients of the corresponding grid cells. The calculation method is to multiply the spectral value within each grid cell by the corresponding terrain-adaptive weight coefficient, so that the spectral data can be... This process effectively reflects the characteristic differences under different terrain conditions, enhancing the correlation between spectral data and the actual evolution environment of cloud clusters. Subsequently, the weighted spectral data is normalized using the z-score normalization method to eliminate dimensional differences between spectral data from different times and channels. The infrared channel spectral values ​​are measured in units of brightness and temperature, while the visible light channel spectral values ​​are measured in units of reflectance. Normalization brings all spectral data to the same numerical order of magnitude. Finally, normalized spectral characteristic data of the target cloud cluster is obtained. This data not only preserves the temporal variation of spectral characteristics but also incorporates information on the impact of terrain differences, improving the data's adaptability to subsequent time-series prediction models.

[0037] Step 308: Based on the normalized spectral feature data of the target cloud cluster, a sliding sampling window for time series prediction is constructed. Specifically, this includes: First, determining the time span of the sliding sampling window based on the temporal length and frequency of change of the normalized spectral feature data of the target cloud cluster. The temporal length is 90 to 120 minutes of data corresponding to 6 to 8 consecutive frames of imagery. The frequency of change is determined by calculating the variance of the difference between adjacent frames of spectral data. Combined with the typical evolution cycle of strong convective clouds, the time span of the sliding sampling window is determined to include 5 to 8 time frames, ensuring that it can completely capture the feature information of a typical change cycle of the cloud cluster. Based on this, and in accordance with the input data format requirements of the time series prediction model, a sliding sampling window for time series prediction is constructed. The starting position of the window is determined by starting from the first frame of the time series data and sliding sequentially backward. The data extraction method is to continuously extract all time series data points within the time span of the window. At the same time, the arrangement format of the data within the window is specified as row vector or column vector to adapt to the input dimension requirements of the model. This window framework enables the subsequent time series sampling process to be more regular, ensuring that the sampled data can accurately adapt to the input requirements of the time series prediction model, laying the foundation for the subsequent generation of high-quality time series samples.

[0038] Step 309: Set a sliding step size smaller than the time span of the sliding sampling window. Based on the sliding step size, perform multiple overlapping time-series sliding window samplings on the normalized spectral feature data to obtain multiple sets of overlapping time-series feature samples. Specifically, this includes: First, based on the sliding sampling window, set a sliding step size smaller than the window time span. Combining the time interval of the time-series images, set the sliding step size to 15 minutes corresponding to one time-series frame. This step size setting ensures that a sufficient number of samples are obtained while also ensuring reasonable differences between adjacent samples, fully capturing the detailed changes in the time-series data. Based on this, use the set 15-minute sliding step size... At intervals, continuous temporal sliding window sampling is performed on the normalized spectral feature data obtained in step 307. During each sampling, the sliding window is fixed at the current position, and all temporal data within the window coverage area is extracted as a temporal feature sample. During the sampling process, it is ensured that the window does not exceed the boundary range of the temporal data. Through multiple overlapping sampling, multiple sets of overlapping temporal feature samples are finally obtained. The overlap between adjacent samples is 4 / 5 to 5 / 6 of the window time span. These samples can comprehensively cover the temporal change details of spectral features, fully characterize the evolution characteristics of cloud clusters in different time segments, and improve the ability of samples to characterize the temporal evolution law of cloud clusters.

[0039] Step 310: Input the multiple sets of overlapping temporal feature samples into the pre-trained ensemble learning prediction model to obtain multiple sub-prediction sequences; perform weighted fusion processing on the multiple sub-prediction sequences to obtain the normalized spectral feature temporal data of the target cloud. Specifically, this includes: First, preparing the pre-trained ensemble learning prediction model. The construction, training, and implementation process of the ensemble learning prediction model is as follows: First, construct the architecture of the sub-prediction models, building three independent sub-prediction models: a Long Short-Term Memory (LSTM) network, a gated recurrent unit (GRU), and a temporal convolutional network. The LSM network is designed with a multi-layer hidden layer structure, capturing long-term dependencies in the temporal data through a gating mechanism; the GRU simplifies the gating structure of the LSM network, improving model training efficiency and real-time performance; the temporal convolutional network adopts a multi-layer causal convolution structure, expanding the receptive field through dilated convolutions to adapt to the multi-scale feature extraction requirements of the temporal data; after completing the sub-model architecture construction, integrate the three sub-models in parallel to build the basic framework of the ensemble learning prediction model, and determine that the fusion method of the sub-model outputs is weighted fusion.

[0040] Following this, model training was conducted. First, a training dataset was prepared, selecting time-series spectral samples of strong convective clouds covering different meteorological conditions, time periods, and intensities. The samples underwent preprocessing, including missing value imputation, outlier removal, and data normalization. Simultaneously, time-series data augmentation techniques were used to expand the sample size, enhancing its diversity and representativeness. The preprocessed dataset was then divided into training, validation, and test sets according to a predetermined ratio. Adaptive training parameters were set for each sub-model, including batch size, initial learning rate, learning rate decay strategy, and maximum number of iterations. A loss function adapted to the time-series prediction task was used to measure the prediction accuracy. The difference between the results and the true values ​​is analyzed. The training set samples are input into three sub-prediction models for independent iterative training. After each preset number of iterations, the model performance is evaluated using a validation set. The model parameters are dynamically adjusted based on the evaluation results until the prediction accuracy and stability of each sub-model on the validation set reach a preset threshold. After the sub-model training is completed, the generalization ability of each sub-model is tested using a test set. Based on the prediction accuracy, stability, and anti-interference ability obtained from the test, the initial weight allocation ratio of each sub-model is determined. The overall training and optimization of the ensemble learning prediction model is completed to ensure that each sub-model has stable time series prediction capabilities.

[0041] After model preparation, multiple sets of overlapping temporal feature samples are input into each sub-prediction model. Each sub-model processes the input samples based on its learned temporal evolution rules and outputs corresponding sub-prediction sequences. Each sub-prediction sequence contains spectral feature prediction values ​​for multiple future times. Based on this, the weights of different sub-prediction sequences are determined according to the prediction accuracy and stability of each sub-model on the validation set. The prediction sequence of the Long Short-Term Memory Network accounts for 40% of the weight, the prediction sequence of the Gated Recurrent Unit accounts for 30%, and the prediction sequence of the Temporal Convolutional Network accounts for 30%. Multiple sub-prediction sequences are fused using a linear weighted summation algorithm. During the fusion process, the prediction values ​​of each sub-prediction sequence are checked for consistency, and abnormal prediction values ​​are removed. Finally, normalized spectral feature temporal data of the target cloud is obtained. This data can combine the prediction advantages of multiple models, reduce the prediction bias of a single model, and improve the reliability and stability of the temporal data.

[0042] Step 311: Based on the normalized spectral feature time-series data of the target cloud cluster, a sliding window method is used to continuously extract fixed-length subsequences in chronological order to obtain a set of time-series data windows arranged by time. Specifically, this includes: First, using the normalized spectral feature time-series data of the target cloud cluster as the processing object, the sliding window time span set in step 308 is followed. If the length of the time-series data changes, the window time span is adjusted appropriately to ensure that the window length can completely represent the state characteristics of the cloud cluster at a certain moment. Usually, the adjusted window length still contains 3 to 5 time-series frames; on this basis, according to time... The process involves continuously sliding the data in a sequence, with each time frame as the sliding step, to extract fixed-length subsequences from the time-series data. Each subsequence corresponds to a specific time segment, and the midpoint of that time segment is the characteristic time of the corresponding subsequence. These extracted subsequences are then arranged in chronological order, with each subsequence serving as an independent time-series data window. A corresponding timestamp is added to each window, resulting in a set of time-arranged time-series data windows. This window sequence can organize continuous time-series data into structured processing units, facilitating the subsequent extraction of cloud state information time-by-time.

[0043] Step 312: For each time-series data window, extract the cloud cluster's intensity value, spatial movement vector, and spectral feature vector at the corresponding time to construct the cloud cluster's state vector at that time. Specifically, this includes: First, for each time-series data window, extract multi-dimensional feature parameters that reflect the core attributes of the cloud cluster. The intensity value, representing the strength of the cloud cluster, is obtained by calculating the minimum and average values ​​of the infrared channel spectral values ​​within the window. The minimum value reflects the intensity of the strongest convection region of the cloud cluster, and the average value reflects the overall intensity of the cloud cluster. The spatial movement vector, reflecting the cloud cluster's direction and velocity, is obtained by calculating the difference between the centroid coordinates of the cloud clusters in adjacent time-series frames within the window. The direction represents the direction of movement, and the ratio of the coordinate difference to the time interval represents the movement speed. The spectral feature vector, which reflects the essential attributes of the cloud, is obtained by extracting the mean, variance, peak value, and other statistical quantities of the visible and infrared spectral values ​​within the window. Based on this, the extracted intensity values, spatial movement vector, and spectral feature vector are fused and integrated. After unifying the dimensions of each parameter, they are concatenated to form a vector, thus constructing the cloud state vector for each time series data window. This state vector can comprehensively characterize the state of the cloud at a specific time from multiple dimensions such as intensity, movement, and spectrum, providing complete input data for the iterative deduction of subsequent time series evolution models.

[0044] Step 313: Input the cloud state vector into a pre-trained temporal evolution model according to the time sequence. The temporal evolution model takes the cloud state vector of the previous moment as input, iteratively calculates and outputs the predicted cloud state value of the next moment. Through continuous iteration, it generates the initial state projection values ​​of the cloud at multiple future moments. Specifically, this includes: First, retrieving the pre-trained temporal evolution model. The construction, training and implementation process of the temporal evolution model is as follows: First, the model architecture is constructed, adopting a sequence-to-sequence model architecture based on an attention mechanism. The model consists of two parts: an encoder and a decoder. The encoder adopts a multi-layer recurrent neural network structure, which is responsible for processing the input historical cloud state vector sequence. The algorithm extracts features to capture temporal dependencies and evolution patterns in the sequence. It also embeds an attention mechanism module to focus on core features that play a key role in future evolution by calculating the attention weights of the state vectors at each time step in the input sequence, thereby improving the targeting of feature extraction. The decoder also adopts a multi-layer recurrent neural network structure. Its input is the high-dimensional temporal features output by the encoder. It is responsible for mapping the high-dimensional features to the cloud state vectors at future time steps. The decoder and encoder are associated through the attention mechanism to ensure that the decoding process can accurately associate key information in the historical sequence. At the same time, a feedback connection mechanism is set up to feed back the decoding output of the previous time step to the current decoding input to ensure the continuity of temporal inference.

[0045] After completing the model architecture construction, model training was carried out. First, a training dataset was prepared, selecting time-series state samples of strong convective clouds covering different meteorological conditions, types, and evolution stages. Each sample contained a continuous historical cloud state vector sequence and corresponding future cloud state vector sequence labels. The samples were preprocessed, including sequence length normalization, missing value imputation, and outlier removal, to ensure the integrity and consistency of the sample data. Simultaneously, time-series augmentation techniques were used to expand the sample size and improve its diversity and representativeness. The preprocessed dataset was then divided into training, validation, and test sets according to a predetermined ratio. Suitable training parameters were set for the model, including training batch size, initial learning rate, learning rate decay strategy, and maximum number of iterations. The model uses a loss function commonly used in time series prediction tasks to measure the difference between the future state vector sequence output by the model and the sample label sequence. Training set samples are input into the model for iterative training. After each iteration, the model's inference performance is evaluated using a validation set. Evaluation metrics include the error between the inference result and the true value, and the consistency of the time series evolution trend. The network parameters and attention weight allocation rules of the model are dynamically adjusted based on the evaluation results. Training stops when the model's evaluation metrics on the validation set reach a preset threshold and tend to stabilize. Subsequently, the model's generalization ability is tested using a test set to verify the model's inference accuracy and stability on unseen cloud time series samples. The final optimization of the model is then completed, ensuring that the model has the ability to accurately predict future states based on historical states.

[0046] After model retrieval is completed, the cloud state vectors are arranged into an input sequence according to time sequence and input into the time-series evolution model. The model takes the cloud state vector of the previous moment as input, focuses on key evolution features through an internal attention mechanism, and calculates and outputs the cloud state prediction value of the next moment in combination with the preset inference rules. The calculation is carried out continuously according to the above iterative logic. The output result of each iteration is used as one of the inputs of the next iteration, and the state prediction values ​​of the cloud at multiple future moments are obtained in sequence. The interval between prediction moments is consistent with the satellite observation interval of 15 minutes, and the prediction duration covers the next 1 to 2 hours. These prediction values ​​together constitute the initial state inference values ​​of the cloud at multiple future moments, providing basic inference data for subsequent spatial correction.

[0047] Step 314: Based on the projected initial state values ​​of the cloud cluster at multiple future moments and the multi-layered grid structure, map the projected initial state value at each moment to the corresponding grid cell, forming a gridded state distribution at each moment. Specifically, this includes: First, obtaining the multi-layered grid structure and determining the spatial coordinates, boundary range, and row and column numbers of each level of grid cell to ensure that the spatial reference of the grid structure is consistent with the spatial reference of the cloud cluster state projection values. Based on this, for the projected initial state values ​​of the cloud cluster at multiple future moments, use a coordinate interpolation-based spatial mapping algorithm to accurately map the projected initial state value at each moment to the corresponding grid cell. On the grid cells, for cases where the coordinates corresponding to the inferred values ​​fall on the grid cell boundaries or between multiple grid cells, a bilinear interpolation method is used to assign the inferred values ​​to the relevant grid cells. This ensures that each grid cell corresponds to a cloud cluster state inference data, including parameters such as cloud cluster intensity, movement speed, and development probability within the cell. The mapping results at each time point are organized, and the inference data are arranged according to the row and column numbers of the grid cells to form a gridded state distribution corresponding to each time point. This distribution can intuitively present the spatial state distribution of the cloud cluster at different future times, providing a structured data carrier for subsequent spatially differentiated terrain correction.

[0048] Step 315: Based on the terrain-adaptive grid weight parameters, spatially weighted correction is performed on the predicted intensity and development probability within each cell of the gridded state distribution to obtain the weighted corrected state parameters of each grid cell. Specifically, this includes: First, calling the terrain-adaptive grid weight parameters obtained in step 305, matching the terrain influence weight corresponding to each grid cell through the row and column numbers of the grid cells to ensure a one-to-one correspondence between the weight parameters and the grid cells of the gridded state distribution; based on this, for the gridded state distribution at each time point, spatially weighted correction is performed on the cloud prediction intensity and development probability within each grid cell. During the correction process... First, the predicted intensity is multiplied by the corresponding terrain influence weighting coefficient to obtain the corrected intensity value. Then, the calculation threshold of the development probability is adjusted by the terrain influence weighting coefficient to make the development probability calculation of areas with greater terrain influence more consistent with the actual terrain conditions. This ensures that the prediction results for areas with greater terrain influence can fully reflect the terrain effect. For example, the intensity correction coefficient of the mountain grid cell is higher than that of the plain area. Through the above correction process, the weighted corrected state parameters of each grid cell are obtained. These parameters include the corrected intensity, movement speed, development probability, etc., which can accurately characterize the cloud state after considering terrain differences and improve the spatial adaptability of the prediction results.

[0049] Step 316: Merge the weighted corrected state parameters of all grid cells to obtain a corrected cloud state evolution sequence covering multiple future moments and including spatial differences. Specifically, this includes: First, collecting the weighted corrected state parameters of all grid cells, labeling each parameter with its corresponding time and spatial coordinates within the grid cell, and determining the spatiotemporal attributes of each parameter; Based on this, using the Kriging interpolation algorithm, fusing the corrected state parameters of different grid cells at the same time, filling the spatial data gaps between grid cells, and generating continuous cloud spatial state distribution raster data. The resolution of this raster data is consistent with the high-resolution grid, ensuring the accuracy of spatial details; Organizing and integrating the cloud spatial state distribution raster data from different moments in chronological order, with each moment's raster data serving as a node in the evolution sequence, with a time interval of 15 minutes, forming a corrected cloud state evolution sequence covering multiple future moments and including spatial differences; This sequence fully presents the spatiotemporal evolution process of the cloud under conditions considering terrain differences, including details such as the cloud's spatial movement trajectory, intensity changes, and expansion or contraction of coverage area, providing comprehensive evolution data support for subsequent extraction of impact information.

[0050] Step 317: Based on the corrected cloud cluster state evolution sequence and combined with the airport's geographical location, extract and obtain the corrected cloud cluster future evolution trend and potential impact information on the airport. Specifically, this includes: First, using the airport geographical location information determined in step 301, determining the latitude and longitude range of the airport core area and related operating areas, including the extended lines at both ends of the runway, the 3-kilometer radius around the terminal building, and the coverage area of ​​the main approach routes; Based on this, using the corrected cloud cluster state evolution sequence obtained in step 316, analyzing the spatial movement trajectory, intensity change trend, and coverage change of the cloud cluster at each time step. The movement trajectory is obtained by connecting the centroid coordinates of the cloud cluster at each time step, and the intensity change trend is obtained by... The changes in the average intensity of the cloud cluster at different times are obtained, and the changes in coverage are obtained by statistically analyzing the number of grid cells covered by the cloud cluster at each time. Combined with the geographical location of the airport, it is determined whether the cloud cluster will cover the airport area in the future, the specific time period of coverage, and the intensity level of the cloud cluster at the time of coverage. At the same time, the duration of the cloud cluster covering the airport is analyzed. Thus, the corrected future evolution trend of the cloud cluster and its potential impact on the airport are extracted and obtained. The evolution trend includes the direction of movement, the speed of movement, and the rate of change of intensity. The potential impact information includes the period of impact, the area of ​​impact, and the intensity level of impact. This information can directly meet the airport's refined early warning needs and provide a core basis for generating targeted early warning instructions.

[0051] In this embodiment of the invention, a polygonal monitoring area is constructed by combining the airport's geographical location and flight route distribution to accurately focus on the potential impact range of cloud clusters on the airport, ensuring the targeted nature of the monitoring. Multiple meshing is used to generate a nested multi-level mesh structure, with different spatial resolutions adapting to different scales of analysis needs. Elevation data of the corresponding areas of the multi-level meshes are extracted to provide a basic data source for subsequent terrain feature analysis. The terrain undulation and aspect characteristics of the mesh units are calculated to quantify the core attributes of the terrain. The terrain influence weight coefficients are normalized to ensure the comparability of weights at different levels of meshes. Multi-level weights are fused to obtain terrain-adaptive mesh weight parameters, integrating multi-scale terrain influence information and improving the applicability of the parameters to subsequent processing.

[0052] Spectral values ​​of the target cloud cluster corresponding to the region are extracted from continuous time-series images to obtain complete original spectral feature time-series information, providing continuous and complete basic data support for subsequent time-series analysis. Spatial weighting and normalization calculations are performed using terrain-adaptive grid weight parameters to incorporate terrain-differentiated influence information into the spectral data. A sliding sampling window for time-series prediction is constructed to provide a framework for standardized sampling of time-series data. Overlapping sampling is performed using a sliding step size smaller than the window time span to increase the number and diversity of time-series feature samples. Multiple sub-prediction sequences are output by an ensemble learning model and then weighted and fused to fully utilize the prediction advantages of multiple models. The reliability and stability of normalized spectral feature time-series data are improved, enhancing the reference value of the time-series data.

[0053] A sliding window is used to extract a fixed-length subsequence to obtain a time-series data window, which standardizes the time-series data format and arranges it in chronological order, facilitating the subsequent extraction and analysis of cloud state vectors. Intensity values, spatial movement vectors, and spectral feature vectors are extracted to construct the cloud state vector, comprehensively characterizing the core attributes of the cloud at a specific moment. The time-series evolution model iteratively calculates and outputs multi-time-step initial state projection values, enabling continuous projection of the cloud's future state. Complete initial data for the future evolution of the cloud is generated, providing a foundation for subsequent spatial correction. The initial state projection values ​​are mapped to multi-layer grid cells, quantifying the abstract cloud state into a concrete form. The system utilizes spatial distribution data to accurately correlate cloud state with spatial location, adapting to subsequent spatial correction processing. It performs spatial weighted correction based on terrain-adaptive grid weight parameters, mitigating the impact of terrain differences on cloud state prediction. It integrates the weighted corrected state parameters of all grid units, consolidating comprehensive spatial information. It generates a complete sequence of corrected cloud state evolution, fully presenting the future spatiotemporal evolution of the cloud. It extracts evolution trends and potential impact information from the airport's geographical location, achieving precise alignment between prediction results and actual airport needs. This provides direct and effective core evidence for generating subsequent early warning instructions, ensuring the application value of the information.

[0054] In a preferred embodiment of the present invention, step 400 above, based on the corrected future evolution trend of the cloud cluster and its potential impact on the airport, combined with the airport's real-time operational status parameters, generates a graded early warning instruction including the impact level, the expected impact period, and recommended handling measures, and issues it in real time, including: Step 401: Based on the corrected future evolution trend of the cloud cluster and its potential impact on the airport, extract the parameters of impact duration, impact intensity, and spatial proximity. Specifically, this includes: First, based on the corrected future evolution trend of the cloud cluster and its potential impact on the airport, determine the spatiotemporal evolution details of the cloud cluster and the impact correlation data contained in the information; on this basis, extract the impact duration parameter. This parameter is determined as follows: by analyzing the future evolution sequence of the cloud cluster, extract the time span of the cloud cluster covering the airport core area and related operating areas, and finally obtain this parameter, that is, starting from the moment the cloud cluster first enters the preset airport impact range. The duration of the impact is calculated from the moment the cloud completely leaves the area until the cloud completely moves away from the airport. Simultaneously, the impact intensity parameter is extracted, combining the average intensity, maximum intensity, and rate of change of the cloud during the airport coverage period in the cloud evolution sequence to comprehensively determine the impact intensity parameter. Furthermore, the spatial proximity parameter is extracted, quantifying the urgency of the cloud's approach to the airport based on the shortest distance between the cloud's future trajectory and the airport's core area, combined with the cloud's movement speed. Through these extraction operations, three key parameters characterizing the core dimensions of the cloud's impact on the airport are obtained, laying the foundation for subsequent decision analysis integrating airport operational status.

[0055] Step 402: Combining the airport's real-time operational status parameters, which include current flight takeoff and landing schedules, runway usage status, and airspace traffic distribution, the real-time operational status parameters are quantified into multi-dimensional operational status indicators. Specifically, this includes: First, collecting the airport's real-time operational status parameters, where the current flight takeoff and landing schedules include the flight departure and landing times, flight numbers, and corresponding runway usage arrangements for the next 2 hours; runway usage status includes the current occupancy status, idle periods, and planned usage duration of each runway; and airspace traffic distribution includes the real-time number of flights and flight altitude levels in the airport's approach airspace and terminal area airspace. Distribution and traffic density; based on this, the above real-time operational status parameters are quantified in multiple dimensions. The current flight take-off and landing plan is quantified into two indicators: take-off and landing flight density per unit time and flight concentration during critical periods. The runway usage status is quantified into two indicators: runway occupancy rate and idle redundancy time. The airspace traffic distribution is quantified into two indicators: traffic saturation and traffic growth rate of each airspace zone. Through the above quantification operations, multi-dimensional operational status indicators are obtained, transforming the originally scattered operational status information into computable and fusionable structured data, providing a data format adaptation basis for subsequent deep integration with cloud cluster impact parameters.

[0056] Step 403 involves fusing the impact duration, impact intensity, and spatial approximation parameters with the multi-dimensional operational status indicators to obtain a multi-dimensional fused feature vector. This multi-dimensional fused feature vector is then mapped to a multi-dimensional joint decision space constructed based on the multi-dimensional operational status indicators. Specifically, this includes: first, fusing the impact duration, impact intensity, and spatial approximation parameters with the multi-dimensional operational status indicators. During the fusion process, all parameters and indicators are first normalized to eliminate the influence of different dimensions. Then, corresponding weights are assigned according to the importance of each parameter and indicator to the early warning decision, and the result is obtained through linear weighted concatenation. A unified multidimensional fusion feature vector is generated. Based on this, a multidimensional joint decision space is constructed. This decision space uses multidimensional operational status indicators as the basic dimension, with each dimension corresponding to the quantification range of an operational status indicator. At the same time, the quantification range of cloud cluster influence parameters is incorporated as an auxiliary dimension to form a multidimensional decision analysis space. Subsequently, the generated multidimensional fusion feature vector is mapped to this multidimensional joint decision space through a spatial coordinate mapping algorithm, so that the fused comprehensive information is transformed into specific coordinate points within the decision space, realizing the systematic integration of scattered information and providing a visualized and standardized analytical basis for subsequent determination of influence level.

[0057] Step 404: Based on the multidimensional joint decision-making space and its predefined impact level labels, locate the multidimensional fused feature vector in the multidimensional joint decision-making space, determine its predefined impact level label according to its spatial location, and obtain the comprehensive impact level; simultaneously, based on the impact duration parameter, determine the expected impact period, specifically including: First, determine the predefined impact level labels within the multidimensional joint decision-making space. The predefinition process of these impact level labels is as follows: First, determine the core basis for the classification, combining the two dimensions of cloud cluster impact degree and airport operational pressure. The cloud cluster impact degree includes parameters such as impact duration, impact intensity, and spatial proximity, while airport operational pressure includes indicators such as flight takeoff and landing density, runway occupancy rate, and airspace traffic saturation; then, determine the number of level divisions. According to civil aviation operation safety management regulations and early warning and response requirements, the impact level is divided into multiple levels, usually divided into... The impact is categorized into four levels, from lowest to highest: minor impact, moderate impact, significant impact, and severe impact. Based on this, specific quantitative judgment criteria are established for each level. By collecting extensive historical data on severe convective weather impacts, the combined characteristics of cloud impact parameters and airport operational status indicators in different cases are statistically analyzed. Combined with the experience and judgment of experts in civil aviation meteorology, air traffic control, and operational support, the parameter threshold range corresponding to each level is determined. The parameter threshold range for each level is mapped to a multi-dimensional joint decision-making space, transforming it into a specific regional range within the decision-making space. Through multiple historical case backtests and parameter optimizations, the boundaries of each regional range are adjusted to ensure that the boundary delineation can accurately distinguish scenarios of different impact levels. Finally, the pre-definition of impact level labels is completed, with each level corresponding to a specific regional range within the decision-making space, and the boundaries of each regional range are determined through extensive historical case data and expert experience verification.

[0058] Based on this, the spatial positioning algorithm is used to determine the specific coordinates of the multi-dimensional fusion feature vector in the decision space, determine the predefined area range to which the coordinates belong, and then determine the corresponding predefined impact level label to obtain the comprehensive impact level. At the same time, based on the extracted impact duration parameter, combined with the current time and the expected time when the cloud first affects the airport, the expected impact period is determined. This period marks the specific time nodes when the cloud begins to affect the airport, when the impact reaches its peak, and when the impact ends, so that the early warning information has clear time attributes.

[0059] Step 405: Based on the comprehensive impact level and the expected impact period, query the preset early warning measure mapping table, match and generate corresponding control suggestions and handling measures, and form a graded early warning instruction. Specifically, this includes: First, pre-constructing the early warning measure mapping table. The pre-construction process of this early warning measure mapping table is as follows: First, determine the core index dimensions of the mapping table, and lock the comprehensive impact level and the expected impact period as the core index. The comprehensive impact level is divided into levels according to the preset minor impact, general impact, major impact, and severe impact. The expected impact period is divided into intervals according to the duration of the impact and whether it covers the peak flight period. Then, collect the core basis for the construction of the mapping table, sort out the relevant regulations on civil aviation operation safety management, airport emergency response plans under severe convective weather, comprehensively summarize historical handling cases under different impact scenarios, and integrate the practical experience of experts in multiple fields such as civil aviation control, meteorological support, and ground operations. Based on the above basis, classify and sort according to the core index dimensions, and for each comprehensive impact level... The system uses a combination index of the impact level and the expected impact period to match corresponding control recommendations and response measures. The control recommendations detail specific aspects such as flight traffic control, runway usage adjustments, and airspace and route optimization. The response measures define the operational procedures for flight delays, diversions, and cancellations, as well as the allocation process, responsible parties, and time requirements for ground support resources. After initial matching, the suitability of the matched content is verified through backtesting using historical cases. The details of the measures are optimized and adjusted in conjunction with expert review, eliminating content lacking practicality and supplementing missing key response steps to ensure that the measures corresponding to each index accurately adapt to the actual impact scenario. After multiple rounds of verification and optimization, a pre-set early warning measure mapping table with a fixed structure and standardized content is formed. This mapping table uses the comprehensive impact level and the expected impact period as the core index. Each index corresponds to a set of specific content containing control recommendations and response measures. The content in the mapping table is determined and continuously optimized based on civil aviation operation safety regulations, historical response cases, and expert consultation.

[0060] Based on this, the comprehensive impact level and the expected impact period are used as search criteria to query the preset early warning measure mapping table and match control recommendations and disposal measures that are highly adapted to the current impact scenario. Subsequently, the comprehensive impact level, the expected impact period, and the matched control recommendations and disposal measures are integrated to form a complete hierarchical early warning instruction.

[0061] Step 406 involves pushing and issuing the tiered early warning instructions in real time through the Civil Aviation Meteorological Information Release System. This includes: First, standardizing the format of the generated tiered early warning instructions by organizing the instruction content according to the standard format of civil aviation meteorological information, determining the release time, scope of application, core information, and execution requirements, ensuring that the instruction format conforms to the information reception specifications of relevant airport departments. Based on this, the standardized tiered early warning instructions are connected to the Civil Aviation Meteorological Information Release System, a standardized information transmission platform dedicated to civil aviation, capable of accurately pushing instructions to relevant units such as airport control centers, airline operation control centers, ground support departments, and air traffic control branches. Through this system, the tiered early warning instructions are pushed to the terminal devices of relevant units in real time and simultaneously publicly released on the system platform, ensuring that all units can simultaneously obtain early warning information, providing information support for timely flight control, ground response, and other response work, and ensuring the timeliness and coverage of the early warning instructions.

[0062] In this embodiment of the invention, core parameters such as duration of impact, intensity of impact, and spatial proximity are extracted, focusing on the key dimensions of cloud clusters' impact on airports. This provides highly targeted basic data for subsequent fusion analysis with airport operational status data, ensuring the relevance and efficiency of subsequent decision analysis. Real-time airport operational status parameters are quantified into multi-dimensional operational status indicators, realizing the transformation of unstructured operational information into structured, computable data. The organic fusion of impact parameters and multi-dimensional operational status indicators is achieved, generating a unified multi-dimensional fusion feature vector. By mapping to a multi-dimensional joint decision space, dispersed meteorological impact information and operational information are integrated into a standardized decision analysis carrier, improving data efficiency. Based on the systematic and logical integration of the system; and using a multi-dimensional joint decision-making space and predefined impact level label positioning feature vectors, the system achieves standardization and normalization in determining the comprehensive impact level; by querying a pre-set early warning measure mapping table to match control recommendations and response measures, the system achieves automation and precision in generating early warning instructions; it ensures that response measures are highly compatible with the comprehensive impact level and the expected impact period, improving the operability and pertinence of early warning instructions; and it leverages the civil aviation meteorological information release system to push and release graded early warning instructions in real time, ensuring the timeliness and coverage of early warning information transmission; and it ensures that airport control, operations support, and other relevant departments simultaneously obtain early warning information, providing information transmission support for rapid response and handling.

[0063] like Figure 2 As shown, embodiments of the present invention also provide an airport strong convective cloud evolution monitoring system based on geostationary satellite multispectral features, comprising: The acquisition module is used to acquire real-time multispectral remote sensing data of airports and preset ranges monitored by geostationary satellites. The intelligent recognition module is used to perform radiometric and geometric calibration on multispectral remote sensing data to obtain calibrated multispectral images. Based on the calibrated multispectral images, visible light reflectance and infrared brightness temperature are extracted and generated into gridded feature data through spatial resampling. Based on the gridded feature data, a fully convolutional neural network integrating spatial pyramid pooling and dense upsampling convolution is used to perform semantic segmentation to identify and locate target cloud clusters with strong convection characteristics. The correction module is used to construct a polygonal monitoring area based on the airport's influence zone for the target cloud cluster and perform multi-grid processing to obtain terrain-adaptive grid weight parameters; based on the terrain-adaptive grid weight parameters, an ensemble learning sliding step prediction method is used to obtain normalized spectral feature time-series data of the target cloud cluster; based on the normalized spectral feature time-series data, a time-series evolution model is constructed using the sliding window method to generate initial state projection values ​​for the future state of the cloud cluster; and the initial state projection values ​​are spatially corrected by combining the terrain-adaptive grid weight parameters to obtain the corrected future evolution trend of the cloud cluster and its potential impact on the airport. The early warning module is used to generate graded early warning instructions, including the impact level, the expected period of impact, and recommended handling measures, based on the corrected future evolution trend of the cloud cluster and its potential impact on the airport, combined with the airport's real-time operational status parameters, and to issue them in real time.

[0064] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0065] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0066] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for monitoring the evolution of strong convective cloud clusters at airports based on geostationary satellite multispectral characteristics, characterized in that, The method includes: Step 100: Obtain multispectral remote sensing data of the airport and the preset range monitored in real time by geostationary satellites; Step 200: Radiometric and geometric calibration are performed on the multispectral remote sensing data to obtain calibrated multispectral images; visible light reflectance and infrared brightness temperature are extracted based on the calibrated multispectral images, and gridded feature data is generated through spatial resampling; based on the gridded feature data, a fully convolutional neural network integrating spatial pyramid pooling and dense upsampling convolution is used to perform semantic segmentation to identify and locate target cloud clusters with strong convection characteristics. Step 300: For the target cloud cluster, a polygonal monitoring area is constructed based on the airport's influence zone and subjected to multi-grid processing to obtain terrain-adaptive grid weight parameters; based on the terrain-adaptive grid weight parameters, an ensemble learning sliding step prediction method is used to obtain normalized spectral feature time-series data of the target cloud cluster; based on the normalized spectral feature time-series data, a time-series evolution model is constructed using the sliding window method to generate initial state projection values ​​for the future state of the cloud cluster; combined with the terrain-adaptive grid weight parameters, the initial state projection values ​​are spatially corrected to obtain the corrected future evolution trend of the cloud cluster and its potential impact on the airport; Step 400: Based on the corrected future evolution trend of the cloud cluster and its potential impact on the airport, combined with the airport's real-time operational status parameters, a graded early warning instruction is obtained, which includes the impact level, the expected period of impact, and recommended handling measures, and is then released in real time.

2. The method for monitoring the evolution of strong convective clouds at airports based on geostationary satellite multispectral characteristics according to claim 1, characterized in that, Step 100 includes: Acquire full-disk Level 1 data products from the multi-channel scanning imager of a geostationary meteorological satellite; From the full-disk primary data product, extract the raw observation counts of the infrared split-window channel and water vapor channel covering the airport and a preset range; Using the radiometric calibration parameters corresponding to each channel, the original observation count values ​​are converted into brightness temperature data to obtain the radiometrically calibrated channel data; Based on the standard Earth coordinate system, the radiometrically calibrated channel data is geometrically corrected and resampled to obtain equal latitude and longitude projection grid data with uniform spatial resolution, which serves as multispectral remote sensing data.

3. The method for monitoring the evolution of strong convective clouds at airports based on geostationary satellite multispectral characteristics according to claim 2, characterized in that, Step 200 includes: Based on the multispectral remote sensing data, reflectance calibration is performed on the visible light band, and brightness-temperature conversion is performed on the infrared split window channel and water vapor channel to obtain radiometrically calibrated multispectral data. Based on the radiometrically calibrated multispectral data, geometric fine correction and reprojection are performed according to a standard geographic reference system to obtain the calibrated multispectral image. Visible reflectance and infrared brightness temperature are extracted from the calibrated multispectral image, and spatial interpolation resampling is performed to generate gridded feature data with uniform spatial resolution. A fully convolutional neural network semantic segmentation model is constructed. The encoder part of the fully convolutional neural network semantic segmentation model integrates a spatial pyramid pooling structure to extract multi-scale contextual features, and the decoder part adopts a dense upsampling convolutional structure to recover spatial details. The fully convolutional neural network semantic segmentation model is trained using sample data containing annotations of strong convective cloud clusters. The gridded feature data is then input into the trained fully convolutional neural network semantic segmentation model to obtain the class probability of each pixel. Based on the category probability of each pixel, connected component analysis and threshold discrimination methods are used to identify and extract target cloud clusters with strong convection characteristics within a preset range, and determine their spatial location and contour information.

4. The method for monitoring the evolution of strong convective clouds at airports based on geostationary satellite multispectral characteristics according to claim 3, characterized in that, Step 300 includes: For the target cloud cluster, a polygonal monitoring area covering the potential impact area is constructed within a preset spatial buffer distance, taking into account the airport's geographical location and flight route distribution. The polygonal monitoring area is subjected to multi-mesh processing, which includes generating a nested multi-level mesh structure based on at least two different spatial resolutions. Obtain the spatial extent definition data of the multi-layer grid structure; based on the spatial extent definition data, extract the elevation data of the corresponding area from the digital elevation model database; Based on the elevation data, the topographic relief and aspect characteristics of each grid cell are calculated; based on the topographic relief and aspect characteristics, and by integrating the spatial distance information between each grid cell and the airport core area, the topographic influence weight coefficient of each grid cell is calculated and generated. The terrain influence weight coefficients of all grid levels are normalized and fused to obtain terrain adaptive grid weight parameters.

5. The method for monitoring the evolution of strong convective clouds at airports based on geostationary satellite multispectral characteristics according to claim 4, characterized in that, Step 300 further includes: Based on the spatial location and contour information of the target cloud cluster, the infrared and visible light channel spectral values ​​within the corresponding spatial range are extracted from the continuous temporal images of the multispectral remote sensing data to obtain the original spectral feature data of the target cloud cluster. Using the terrain-adaptive grid weight parameters, the original spectral feature data of the target cloud are spatially weighted and normalized to obtain the normalized spectral feature data of the target cloud. Based on the normalized spectral feature data of the target cloud, a sliding sampling window for time series prediction is constructed. Set a sliding step size smaller than the time span of the sliding sampling window, and perform multiple overlapping time-series sliding window samplings on the normalized spectral feature data based on the sliding step size to obtain multiple sets of overlapping time-series feature samples. The multiple sets of overlapping temporal feature samples are input into a pre-trained ensemble learning prediction model to obtain multiple sub-prediction sequences; the multiple sub-prediction sequences are weighted and fused to obtain the normalized spectral feature temporal data of the target cloud.

6. The method for monitoring the evolution of strong convective clouds at airports based on geostationary satellite multispectral characteristics according to claim 5, characterized in that, Step 300 further includes: Based on the normalized spectral characteristics of the target cloud cluster, a sliding window method is used to continuously extract fixed-length subsequences in chronological order to obtain a set of time-series data windows arranged in time. For each time series data window, the intensity value, spatial movement vector and spectral feature vector of the cloud at the corresponding time are extracted to form the cloud state vector at that time. The cloud state vector is input into a pre-trained temporal evolution model in a time series manner. The temporal evolution model takes the cloud state vector of the previous moment as input, iteratively calculates and outputs the cloud state prediction value of the next moment. Through continuous iteration, the initial state projection values ​​of the cloud at multiple future moments are generated. Based on the initial state projection values ​​of the cloud cluster at multiple future moments, and the multi-layer grid structure, the initial state projection values ​​at each moment are mapped to the corresponding grid cells to form the gridded state distribution at each moment. Based on the terrain-adaptive grid weight parameters, the predicted intensity and development probability in each cell of the gridded state distribution are spatially weighted and corrected to obtain the weighted corrected state parameters of each grid cell. By integrating the weighted corrected state parameters of all grid cells, a corrected cloud state evolution sequence covering multiple future time points and including spatial differences is obtained. Based on the corrected cloud cluster state evolution sequence and combined with the airport's geographical location, the corrected cloud cluster future evolution trend and potential impact on the airport are extracted and obtained.

7. The method for monitoring the evolution of strong convective clouds at airports based on geostationary satellite multispectral characteristics according to claim 6, characterized in that, Step 400 includes: Based on the corrected cloud cluster future evolution trend and potential impact on the airport, the impact duration, impact intensity and spatial proximity parameters are extracted. Combining the airport's real-time operational status parameters, including current flight takeoff and landing plans, runway usage status, and airspace traffic distribution, the real-time operational status parameters are quantified into multi-dimensional operational status indicators. The duration of the impact, the intensity of the impact, and the spatial approximation parameters are fused with the multi-dimensional operational status indicators to obtain a multi-dimensional fused feature vector, and the multi-dimensional fused feature vector is mapped to a multi-dimensional joint decision space constructed based on the multi-dimensional operational status indicators. Based on the multidimensional joint decision space and its predefined impact level labels, the multidimensional fusion feature vector is located in the multidimensional joint decision space, and its predefined impact level label is determined according to its spatial location to obtain the comprehensive impact level; at the same time, the expected impact period is determined based on the impact duration parameter. Based on the comprehensive impact level and the expected impact period, a preset early warning measure mapping table is queried, and corresponding control suggestions and disposal measures are generated to form a graded early warning instruction; The tiered early warning instructions will be pushed and released in real time through the civil aviation meteorological information release system.

8. A monitoring system for the evolution of strong convective clouds at airports based on geostationary satellite multispectral characteristics, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The acquisition module is used to acquire real-time multispectral remote sensing data of airports and preset ranges monitored by geostationary satellites. The intelligent recognition module is used to perform radiometric and geometric calibration on multispectral remote sensing data to obtain calibrated multispectral images. Based on the calibrated multispectral images, visible light reflectance and infrared brightness temperature are extracted and generated into gridded feature data through spatial resampling. Based on the gridded feature data, a fully convolutional neural network integrating spatial pyramid pooling and dense upsampling convolution is used to perform semantic segmentation to identify and locate target cloud clusters with strong convection characteristics. The correction module is used to construct a polygonal monitoring area based on the airport influence zone for the target cloud cluster and perform multi-grid processing to obtain terrain adaptive grid weight parameters. Based on terrain-adaptive grid weight parameters, an ensemble learning sliding step size prediction method is used to obtain the normalized spectral feature time series data of the target cloud cluster. Based on normalized spectral feature time series data, a time series evolution model is constructed using the sliding window method to generate initial state projection values ​​for the future state of cloud clusters. By combining terrain-adaptive grid weight parameters, the initial state projection values ​​are spatially corrected to obtain the corrected cloud cluster future evolution trend and potential impact information on the airport. The early warning module is used to generate graded early warning instructions, including the impact level, the expected period of impact, and recommended handling measures, based on the corrected future evolution trend of the cloud cluster and its potential impact on the airport, combined with the airport's real-time operational status parameters, and to issue them in real time.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images