Satellite remote sensing data-based strong convective cloud cluster identification method and system

Abstract

This application discloses a method and system for identifying severe convective clouds based on satellite remote sensing data, relating to the fields of meteorological remote sensing and severe weather monitoring. The method includes: spatiotemporal registration of multi-spectral satellite data to form a multi-channel time-series sequence, introducing matching environmental parameters, and calculating an adaptive dynamic threshold to stably extract candidate cloud objects; fusing multi-spectral visual features and environmental physical features to construct a convective complexity index and generate node features; further, constructing a graph structure with candidate cloud objects as nodes and spatial proximity as edges, inputting it into a physically constrained graph convolutional model, and using environmental physical features for gating constraints in neighborhood aggregation to output the severe convection probability, severity score, and geographical extent of each cloud object. Thus, by employing a fusion mechanism of environmental physical prior and graph structure reasoning, the model's adaptability and output consistency under different climatic backgrounds are improved, and the overall reliability and interpretability of the identification results are enhanced.

Description

Technical Field

[0001] This application relates to the field of meteorological remote sensing and severe weather monitoring technology, and in particular to a method and system for identifying strong convective cloud clusters based on satellite remote sensing data. Background Technology

[0002] Severe convective clouds are a direct manifestation of severe weather events such as tornadoes, torrential rains, and hail. Utilizing satellite remote sensing data for all-weather, all-range monitoring is crucial for disaster prevention and mitigation. Currently, identification primarily relies on visible light, infrared, and water vapor band data provided by meteorological satellites. However, existing identification technologies mainly fall into two categories: traditional physical thresholding methods and emerging deep learning methods. Both of these methods have significant limitations in practical applications, making it difficult to meet the monitoring requirements of high accuracy and high generalization capability.

[0003] Traditional identification methods primarily rely on infrared window region channel brightness temperature thresholding (IRW) or brightness temperature difference (BTD) methods. These methods typically assume extremely low temperatures at the top of strong convective clouds and distinguish targets by setting fixed brightness temperature values ​​or channel differences. Their core drawback lies in the difficulty of adapting to varying geographical and seasonal differences using "fixed thresholds." The algorithms are extremely sensitive to threshold changes and are prone to misclassifying thick cumulus or cirrus clouds as strong convective clouds, resulting in a high false alarm rate. Although some improved methods have introduced tropopause differential methods, they are still limited by operational complexity and the difficulty in standardizing thresholds.

[0004] In recent years, although machine learning and deep learning technologies have been introduced into the field of cloud identification, improving the accuracy of segmentation and classification through models such as convolutional neural networks, these methods are mostly purely data-driven "black box" models. They generally neglect the internal physical characteristics of strong convective systems (such as energy structure and thermodynamic constraints), resulting in poor model interpretability. In addition, deep learning models usually require massive amounts of labeled data, and when faced with new regions or new climatic backgrounds outside the training data, their generalization ability often declines significantly due to the lack of guidance from physical mechanisms. Summary of the Invention

[0005] This application provides a method, system, storage medium, computer program product, and electronic device for identifying strong convective cloud clusters based on satellite remote sensing data, in order to at least solve the problem that the accuracy of strong convective cloud cluster identification is unstable and prone to false alarms and missed alarms in different regions and climate backgrounds in the current related technologies.

[0006] In a first aspect, embodiments of this application provide a method for identifying strong convective cloud clusters based on satellite remote sensing data. The method includes: acquiring multi-spectral satellite remote sensing observation data of a monitored area within at least one observation period, and acquiring environmental parameter data corresponding to the multi-spectral satellite remote sensing observation data in time and space; performing spatiotemporal registration preprocessing on the multi-spectral satellite remote sensing observation data to generate a multi-channel observation data sequence arranged in chronological order; determining an adaptive dynamic threshold for cloud cluster segmentation based on the environmental parameter data, and using the adaptive dynamic threshold to perform candidate extraction processing on cloud areas at the target observation time in the multi-channel observation data sequence to obtain a set of candidate cloud cluster regions, wherein each candidate cloud cluster region in the set of candidate cloud cluster regions corresponds to a set of spatially connected pixel regions; extracting image visual features based on the multi-channel observation data sequence for each candidate cloud cluster region in the set of candidate cloud cluster regions, and determining environmental physical features based on the environmental parameter data, and determining environmental physical features according to the image visual features and the environmental physical features. Feature construction is used to characterize the degree of strong convection in the candidate cloud cluster region, and node feature vectors of the candidate cloud cluster region are generated. The node feature vectors include the image visual features, the environmental physical features, and the convection complexity index. A candidate cloud cluster graph structure is constructed based on the candidate cloud cluster region set. The graph nodes of the candidate cloud cluster graph structure correspond to each candidate cloud cluster region, and the graph edges are established between candidate cloud cluster regions that meet preset distance conditions to characterize the spatial proximity relationship of the candidate cloud cluster regions. The candidate cloud cluster graph structure and the node feature vectors of each graph node are input into a physically constrained graph convolutional integral class model to output the strong convection probability and severity score corresponding to each candidate cloud cluster region. Based on the strong convection probability and the severity score, a strong convection cloud cluster identification result is output. The strong convection cloud cluster identification result includes at least the geographical range of the strong convection cloud cluster and the corresponding severity score. In the graph convolutional integral class model, the neighborhood feature aggregation is constrained based on the environmental physical features during the graph convolutional inference process.

[0007] Secondly, embodiments of this application provide a strong convective cloud cluster identification system based on satellite remote sensing data. The system includes: a data acquisition unit, configured to acquire multi-spectral satellite remote sensing observation data of the area to be monitored within at least one observation period, and acquire environmental parameter data corresponding to the multi-spectral satellite remote sensing observation data in time and space; perform spatiotemporal registration preprocessing on the multi-spectral satellite remote sensing observation data to generate a multi-channel observation data sequence arranged in chronological order; an adaptive candidate extraction unit, configured to determine an adaptive dynamic threshold for cloud cluster segmentation based on the environmental parameter data, and perform candidate extraction processing on the cloud area at the target observation time in the multi-channel observation data sequence using the adaptive dynamic threshold to obtain a candidate cloud cluster region set, wherein each candidate cloud cluster region in the candidate cloud cluster region set corresponds to a set of spatially connected pixel regions; and a physical guided feature construction unit, configured to extract image visual features based on the multi-channel observation data sequence for each candidate cloud cluster region in the candidate cloud cluster region set, and determine environmental physical features based on the environmental parameter data, and construct the image based on the image visual features. A convection complexity index is constructed based on the environmental physical features to characterize the degree of strong convection in the candidate cloud cluster region, and node feature vectors of the candidate cloud cluster region are generated. The node feature vectors include the image visual features, the environmental physical features, and the convection complexity index. A graph structure construction unit is used to construct a candidate cloud cluster graph structure based on the set of candidate cloud cluster regions. The graph nodes of the candidate cloud cluster graph structure correspond to each candidate cloud cluster region, and the graph edges are established between candidate cloud cluster regions that meet preset distance conditions to characterize the spatial proximity relationship of the candidate cloud cluster regions. A physical constraint inference unit is used to input the candidate cloud cluster graph structure and the node feature vectors of each graph node into a physical constraint graph convolutional inference model to output the strong convection probability and severity score corresponding to each candidate cloud cluster region, and output the strong convection cloud cluster identification result based on the strong convection probability and the severity score. The strong convection cloud cluster identification result includes at least the geographical range of the strong convection cloud cluster and the corresponding severity score. The physical constraint graph convolutional inference model constrains the aggregation of neighborhood features based on the environmental physical features during the graph convolutional inference process.

[0008] Thirdly, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the strong convective cloud cluster identification method based on satellite remote sensing data according to any embodiment of this application.

[0009] Fourthly, embodiments of this application provide a storage medium storing a computer program thereon, characterized in that, when the program is executed by a processor, it implements the steps of the strong convective cloud cluster identification method based on satellite remote sensing data according to any embodiment of this application.

[0010] Fifthly, embodiments of this application provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the strong convective cloud cluster identification method based on satellite remote sensing data according to any embodiment of this application.

[0011] The method and system for identifying strong convective cloud clusters based on satellite remote sensing data provided in this application can achieve at least the following technical effects:

[0012] (1) By introducing environmental parameters that correspond to satellite observations in time and space, an adaptive discrimination benchmark driven by "environmental state" is formed, so that the extraction of candidate cloud areas no longer depends on a single fixed value, but is dynamically adjusted with changes in atmospheric background. Since the threshold is established in correspondence with the thermal and stratification conditions at the time and place, the extraction of candidate areas can maintain a relatively stable discrimination scale under different regional, seasonal and climatic backgrounds, thereby improving the effectiveness and consistency of the candidate set from the source. At the same time, the spatiotemporal alignment of multi-spectral observations ensures that the observation correspondence of the same target in different channels and at different times is stable, reducing the false differences caused by observation mismatch, and making subsequent discrimination based on multi-channel information based on a more consistent observation basis.

[0013] (2) At the candidate region level, a node feature vector integrating multi-spectral visual representation, environmental physical representation, and convection complexity index is constructed, and a candidate cloud map structure is constructed based on spatial proximity, so that the discrimination criteria are expanded from a single channel or a single index to a comprehensive discrimination of "region-environment-neighborhood context". Thus, when the model is reasoning, it can not only utilize the multi-channel response differences within the candidate region, but also use neighborhood information to suppress unstable decisions caused by isolated anomalies; at the same time, the convection complexity index, as an explicit representation of the degree of strong convection, is incorporated into the node features, so that the output strong convection probability and severity score can distinguish the risk level of different candidate regions in a more granular way, and improve the usability of the recognition results in terms of spatial continuity and severity classification.

[0014] This technical solution couples the spatially structured representation of candidate clouds with environmental physical constraints into graph convolutional inference. During the graph convolutional inference process, a neighborhood aggregation constraint mechanism based on environmental physical features is introduced. This transforms neighborhood information fusion from indiscriminate propagation into a joint assessment of the probability and severity of strong convection under the constraints of spatial neighborhood collaboration and physical consistency. This establishes a more stable coupling relationship between the discrimination process and the background environmental conditions. Therefore, this constraint mechanism improves the model's adaptability and output consistency under different climatic backgrounds, and makes the probability and severity output traceable to the influence of environmental physical characteristics on the aggregation path, thus enhancing the overall reliability and interpretability of the identification results. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating an example of a method for identifying strong convective cloud clusters based on satellite remote sensing data according to an embodiment of this application is shown.

[0017] Figure 2 A schematic diagram of a true-color composite image of the area to be monitored at the target observation time is shown;

[0018] Figure 3 A schematic diagram is shown showing the annotation of the strong convective cloud identification results output by the method of this application superimposed on a true-color composite image;

[0019] Figure 4 It shows the relationship with Figure 2 A schematic diagram of the long-wave infrared brightness temperature map of the same region at the same time;

[0020] Figure 5 A schematic diagram is shown showing the labeling of the strong convective cloud identification results output by the method of this application superimposed on a long-wave infrared brightness temperature map;

[0021] Figure 6 The comparison results of detection accuracy of different methods when brightness temperature threshold changes are shown;

[0022] Figure 7 The heatmap shows the distribution of the Convection Complexity Index (CCI) over time for several candidate cloud clusters selected from the experimental test set.

[0023] Figure 8 A structural block diagram of an example of a strong convective cloud cluster identification system based on satellite remote sensing data according to an embodiment of this application is shown. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] It should be noted that modern meteorological operations have a significant need for "all-day, wide-coverage, and timely" monitoring of severe convective cloud clusters. With the continuous observation capabilities of high-resolution meteorological satellites improving, multi-spectral remote sensing data, including visible light, infrared, and water vapor, can provide multi-dimensional information such as cloud top temperature, cloud structure, and water vapor transport, making continuous monitoring and automated analysis of cloud fields possible. In contrast, traditional manual identification methods rely more on experience-based interpretation and local observation data (such as local radar echoes) for auxiliary judgment, making it difficult to maintain consistency and real-time performance in large-scale, massive data stream scenarios. Therefore, operational teams are gradually shifting towards automatic identification technologies based on satellite remote sensing.

[0026] In the evolution of related technologies, early automatic identification methods mostly started from the brightness temperature of the infrared window channel, distinguishing deep convective clouds from background clouds based on the statistical law that "strong convective cloud top temperatures are lower". To improve identification accuracy and adaptability to complex cloud fields, some studies have further introduced multi-channel brightness temperature differential features, such as using the brightness temperature difference between the water vapor channel and the infrared window channel, or using the difference between cloud top brightness temperature and tropopause temperature to enhance sensitivity to strong convective clouds. Such approaches can enhance the response to specific types of convective structures to a certain extent, but they often face two practical constraints in operational applications: First, multi-channel differential features are greatly affected by factors such as atmospheric water vapor transport and stratospheric disturbances, and are prone to triggering false alarms in non-strong convective scenarios; second, tropopause correlation methods usually rely on additional environmental field data support, and the threshold settings are strongly correlated with regional climate background, making consistent deployment across regions and seasons difficult.

[0027] To reduce false alarms caused by single thresholds or single differential criteria, some studies have attempted to introduce richer "morphology-texture-structure" information. For example, they combine cloud top texture roughness, cloud boundary gradient, and anvil-shaped cloud morphology features for comprehensive discrimination, or employ multi-step scoring and probabilistic output to improve stability. These methods can suppress local noise and false detections to some extent, but their practical application still has limitations: on the one hand, multi-feature scoring systems usually introduce more empirical rules and manually set parameters, which increases the difficulty of parameter transfer and unified calibration when facing differences in cloud field morphology in different regions; on the other hand, in continuous monitoring tasks with high frequency and massive amounts of satellite data, complex multi-step discrimination and post-processing chains can easily lead to additional computational overhead and may result in unstable output due to noise accumulation and error propagation.

[0028] In recent years, machine learning and deep learning methods have been used for cloud segmentation and strong convection identification. Models such as convolutional networks are used to automatically learn discrimination patterns from multispectral remote sensing images, and lightweight structures or model ensembles are employed to balance inference speed. Meanwhile, research has explored incorporating multi-source data (such as radar reflectivity features or environmental parameters from numerical models / reanalysis data) to improve the timeliness and accuracy of identification. Although these methods can achieve high identification accuracy under specific datasets and regional conditions, typical pain points remain in practical applications: First, data-driven models often rely heavily on training distribution, and their performance can fluctuate when the region, season, observation geometry, or cloud field structure changes. Second, the model's decision-making mechanism is difficult to align with the thermodynamic / dynamic constraints of the convective system, leading to insufficient interpretability and difficulty in supporting the operational need for traceability regarding "why an alert was triggered and why the severity was determined." Third, obtaining large-scale labeled samples is costly, and the labeling standards of different operational units may be inconsistent, further limiting the model's stable generalization.

[0029] In addition, there are some common engineering challenges in actual business processes: strong convective cloud clusters often have spatiotemporal evolution characteristics such as rapid development, movement and merging and splitting, and static discrimination at a single moment is easily affected by instantaneous observation noise, thin cloud occlusion or local missing measurements; at the same time, there are correlations between cloud clusters in terms of spatial proximity and development and evolution, and it is difficult to stably characterize the structural information of "mutual influence and co-evolution" in the cloud field by classifying only a single candidate region independently.

[0030] It should be understood that the above description of the relevant technologies is intended only to help the public better understand the inventive spirit and motivation of this application, and is not intended to limit this application. Furthermore, the technical solutions described in the above-mentioned relevant technologies are not prior art, and may also be undisclosed technical solutions, such as those under research or in the laboratory stage.

[0031] The technical solutions in this application, including the collection, storage, use, processing, transmission, provision, and disclosure of users' personal information, comply with relevant laws and regulations and do not violate public order and good morals.

[0032] Figure 1 A flowchart illustrating an example of a method for identifying strong convective cloud clusters based on satellite remote sensing data according to an embodiment of this application is shown.

[0033] Regarding the execution subject of the method in the embodiments of this application, it can be any controller or processor with computing or processing capabilities, such as a cloud-based remote sensing intelligent analysis platform controller. By performing unified data alignment, candidate region generation, node feature vector construction, and physical constraint graph convolutional inference processing on multi-spectral satellite remote sensing observation data and corresponding environmental parameter data, and further generating strong convection probability and severity scores and their corresponding geographical range outputs, the automatic and continuous generation of strong convective cloud cluster identification results for the area to be monitored is realized.

[0034] In some examples, it may be integrated into an electronic device or terminal through software, hardware, or a combination of both, and the type of terminal or electronic device may be diverse.

[0035] like Figure 1 As shown, in step S110, multi-spectral satellite remote sensing observation data of the area to be monitored within at least one observation period is acquired, and environmental parameter data corresponding to the multi-spectral satellite remote sensing observation data in time and space are acquired; spatiotemporal registration preprocessing is performed on the multi-spectral satellite remote sensing observation data to generate a multi-channel observation data sequence arranged in chronological order.

[0036] In some implementations, the multi-spectral satellite remote sensing data may come from multi-channel imaging products of geostationary meteorological satellites or polar-orbiting meteorological satellites, including at least two or three of the visible light channel, infrared window channel, and water vapor channel, so as to simultaneously characterize the top temperature information, thickness information, and upper-level water vapor distribution information of the cloud body.

[0037] The observation period can adopt the fixed update cycle of the operational system (e.g., every 5 or 10 minutes). The area to be monitored can be determined by administrative region, watershed grid, or a pre-selected latitude and longitude area. The environmental parameter data corresponding to the observation data in time and space can come from numerical weather prediction grid fields, reanalysis products, or surface / sounding fusion products. The environmental parameter data can include at least one or more of the following: temperature and humidity profile parameters characterizing the convective environment, convective available potential energy parameters, lifting index parameters, tropopause-related parameters, and vertical wind shear parameters.

[0038] Furthermore, when performing spatiotemporal registration preprocessing on the multi-spectral satellite remote sensing observation data, the following steps may be taken in sequence: radiometric calibration and brightness temperature inversion (for infrared / water vapor channels), geolocation and projection unification, resolution unification (upsampling or downsampling is performed on different channels), missing pixel filling or masking, and cloud top parallax correction (when the business scenario requires more accurate geolocation, parallax correction can be performed on high cloud pixels based on cloud top height estimation).

[0039] After the above processing is completed, the multi-channel gratings at each observation time are stacked in chronological order to form the multi-channel observation data sequence. This ensures that data from different spectral bands and different sources have consistent pixel alignment on the same coordinate system and time axis, improving the reproducibility and stability of the entire identification chain.

[0040] In step S120, an adaptive dynamic threshold for cloud segmentation is determined based on environmental parameter data, and the adaptive dynamic threshold is used to perform candidate extraction processing on the cloud area at the target observation time in the multi-channel observation data sequence to obtain a set of candidate cloud areas. Each candidate cloud area in the set of candidate cloud areas corresponds to a set of spatially connected pixel areas.

[0041] In some embodiments, the adaptive dynamic threshold used for cloud segmentation is not a fixed constant, but is determined by a threshold function or threshold lookup table driven by environmental parameter data. For example, a dynamic threshold can be constructed for the brightness temperature of the infrared window region, which is functionally correlated with tropopause temperature, vertical temperature lapse rate, convective potential energy, precipitable water, or humidity profile characteristics. Furthermore, a dynamic threshold can also be constructed for brightness-temperature difference indicators, adaptively adjusting to changes in upper-level humidity background or tropopause height. In addition, the threshold function can further incorporate a spatial partitioning mechanism, such as dividing the monitored area into several sub-regions or partitioning by pixel topographic height zones, calculating dynamic thresholds separately for different partitions, thereby preventing the threshold from being skewed by a few extreme environmental parameters over a large area.

[0042] After obtaining the dynamic threshold, when performing candidate extraction processing on the cloud region at the target observation time, the multi-channel data at the target observation time can be substituted into the above dynamic threshold rules to generate a binary cloud region mask, and morphological processing can be performed to obtain connected candidate regions. For example, opening operations are first performed to remove isolated noise points, then closing operations are performed to fill small holes, and then the set of spatially connected pixels is extracted through connected component labels (4-connected or 8-connected), and geometric properties such as area, equivalent diameter, and boundary complexity are calculated for each connected component to filter out objects with too small an area or abnormal shape (such as fragments smaller than a few pixels).

[0043] Finally, a set of candidate cloud regions is output, with each candidate region corresponding to a set of spatially connected pixels and its bounding rectangle, outline, and polygon boundaries preserved. Through the above design, the generation of the candidate set can automatically adjust the "selection scale" when the environmental background changes, and output region objects with clear geometric boundaries in the form of connected components. It can maintain a relatively consistent candidate selection scale under different regional / seasonal backgrounds, while the connected components and morphological constraints make the candidate objects closer to real cloud entities.

[0044] In this paper, "environmental background" refers to a set of physical conditions consisting of a set of environmental parameters that correspond to the candidate cloud region in time and space and characterize its large-scale and mesoscale atmospheric state. These parameters include thermal instability, tropopause constraint, temperature and humidity profiles and water vapor content, wind field and wind shear, etc. It is used to characterize whether the region has external physical support for the occurrence and maintenance of convection, and whether such support conditions are consistent among neighboring regions.

[0045] In step S130, for each candidate cloud region in the candidate cloud region set, image visual features are extracted based on multi-channel observation data sequences, and environmental physical features are determined based on environmental parameter data. A convection complexity index is constructed to characterize the degree of strong convection in the candidate cloud region based on the image visual features and environmental physical features, and a node feature vector of the candidate cloud region is generated. The node feature vector includes image visual features, environmental physical features, and convection complexity index.

[0046] In some implementations, for each candidate cloud region, visual features characterizing cloud structure and evolution can be extracted from multi-channel time-series observation sequences. These visual features can include regional statistics, as well as texture and temporal variations. Examples include the mean, minimum, quantiles (e.g., the 5th percentile to highlight the coldest cloud top), and standard deviation (characterizing internal inhomogeneity) of brightness temperature in each infrared / water vapor channel; the mean and extreme values ​​of water vapor-window brightness temperature difference, split window difference, etc. (characterizing cloud top humidity / microphysical differences); texture contrast, entropy, and homogeneity based on the Gray-Level Co-occurrence Matrix (GLCM) or Local Binary Pattern (LBP) (characterizing differences between anvil cloud boundaries and convective core structures); and sequence-based trend features, such as the rate of change of the coldest regional brightness temperature over time (cloud-top cooling rate), regional area expansion rate, and the percentage of cold cloud cores over time, used to reflect the development / maturation stage of convection. To ensure comparability across different scales, the above features can be normalized (e.g., by performing min-max or z-score based on the channel statistical range), and mask statistics can be applied to missing pixels to avoid bias caused by missing measurements.

[0047] Environmental physical characteristics consist of environmental parameters corresponding to the location and time of the candidate region. They can be extracted based on the centroid location of the candidate region, or averaged / quantile converged on the environmental parameters corresponding to the pixels within the region (to improve stability). Typical physical characteristics can be diverse, such as tropopause temperature / height, CAPE (Convective Available Potential Energy) or CIN (Convective Inhibition), total precipitable water, relative humidity of key layers (such as 700 / 500 / 300 hPa), 0-6 km wind shear, and 0°C layer height.

[0048] Furthermore, a convection complexity index can be constructed based on visual and environmental physical features to couple "observed strong convection patterns" with "background supportability" into a single interpretable quantity. Specifically, a weighted combination or constrained mapping is constructed, for example, using "the degree of cloud top cooling / coldest brightness temperature surpassing the tropopause," "water vapor difference intensity," "texture / boundary complexity," and "environmental instability and deep water vapor" as positive components, and "strong suppression (such as CIN) or environmental dry layer indicators" as negative components, forming a normalized index CCI. Finally, visual features, environmental physical features, and CCI are concatenated to form a node feature vector, so that the candidate region not only includes "visible cloud body performance" but also "alignable environmental background," and the coupling relationship between the two is made explicit by the complexity index, providing more sufficient and traceable discrimination basis for subsequent classification models, and making the severity score have interpretable feature support.

[0049] In step S140, a candidate cloud cluster graph structure is constructed based on the candidate cloud cluster region set. The graph nodes of the candidate cloud cluster graph structure correspond to each candidate cloud cluster region, and the graph edges are established between candidate cloud cluster regions that meet the preset distance conditions, which are used to characterize the spatial proximity relationship of the candidate cloud cluster regions.

[0050] In some implementations, the candidate cloud cluster graph structure can be constructed using a consistent "node-edge-weight" definition. Preset distance conditions can be set to a fixed geographical distance threshold (e.g., a threshold in kilometers) according to business needs, or a k-nearest neighbor strategy can be used (i.e., connecting each node to the k nearest candidate regions) to avoid the graph structure becoming too sparse or too dense when the number of candidates changes. Distance calculation can be based on the spherical distance between the centroids of candidate regions or on the minimum distance between the boundaries of candidate regions. When candidate regions overlap or are close to each other, the minimum boundary distance can be prioritized as the proximity criterion to enhance the ability to characterize the "cloud cluster" structure.

[0051] Furthermore, to enhance the graph structure's ability to express spatial organization, graph edges can carry edge attributes or edge weights. For example, edge weights can be defined as distance decay functions, or similarity in brightness temperature, texture, or complexity metrics can be introduced for candidate regions, enabling stronger information interaction between more similar and neighboring candidate regions during graph convolutional inference. Through the above graph structure modeling, the model can explicitly utilize spatial context at the region level. During inference, the model no longer treats each candidate cloud cluster in isolation, but explicitly utilizes the common spatial organization and neighborhood consistency of cloud systems, reducing sensitivity to local noise or isolated anomalies and improving the overall recognition coherence of complex scenarios such as "multi-monosystem systems" and "zonal convection systems."

[0052] In step S150, the candidate cloud cluster graph structure and the node feature vectors of each graph node are input into the physical constraint graph convolutional integrator model to output the strong convection probability and severity score corresponding to each candidate cloud cluster region, and output the strong convection cloud cluster identification result based on the strong convection probability and severity score. The strong convection cloud cluster identification result includes at least the geographical range of the strong convection cloud cluster and the corresponding severity score. In the graph convolutional integrator model, the physical constraint graph convolutional integrator model constrains the aggregation of neighborhood features based on environmental physical features during the graph convolutional inference process.

[0053] In some implementations, the physically constrained graph convolutional integral class model may include a graph convolutional feature extraction layer and an output prediction layer. The graph convolutional feature extraction layer is used to perform neighborhood feature aggregation on the candidate cloud graph structure, and the output prediction layer is used to map the updated node representations to strong convection probabilities and severity scores.

[0054] More specifically, to constrain neighborhood feature aggregation based on environmental physical characteristics, an environmental consistency gating mechanism can be introduced during the calculation of neighborhood aggregation weights. For example, for nodes... Its neighboring nodes The physical consistency coefficient is calculated based on the differences in the physical feature vectors of the two environments (such as differences in instability, tropopause constraints, and humidity background). This coefficient is then used as a modulatory term for the aggregation weights, suppressing the aggregation contribution of adjacent nodes with significantly inconsistent environmental backgrounds, while enhancing the aggregation contribution of nodes with consistent environmental backgrounds and spatial proximity. Thus, the model can follow the constraint logic that "spatial proximity regions in similar environments should share more discriminative information" during inference, thereby reducing the interference caused by unreasonable cross-background information propagation on the output.

[0055] In terms of output and post-processing, the probability of severe convection can be obtained by the classification head and normalized by the activation function. For example, the severity score can be output by the regression head or the hierarchical classification head to characterize the risk level or severity of severe convection. Subsequently, target candidate regions can be screened based on the probability threshold of severe convection, and multiple high-probability nodes can be merged into severe convective cloud objects according to spatial connectivity or graph connectivity rules. For each object, its geographical extent can be calculated (e.g., generating polygonal boundaries or raster envelopes from a set of pixels), and an object-level severity score can be generated based on the maximum, mean, or area-weighted severity scores within the region.

[0056] Through the above reasoning and output mechanism, the business system can simultaneously obtain structured identification results of "where" (geographical range) and "how strong" (severity) of strong convective cloud clusters. Furthermore, because the aggregation process is modulated by environmental physical constraints, the output has a more stable and consistent performance across regions and background scenarios, which facilitates the formation of a sustainable monitoring product.

[0057] For example, after inference output, the probability of strong convection can be filtered according to business thresholds (e.g., candidate regions with probabilities exceeding a set threshold are identified as results), and severity scores are combined to achieve classification; in spatial product generation, candidate connected regions that meet the conditions can be directly output as pixel boundaries or vector polygons, while adjacent and highly confident candidate regions are merged to obtain the geographical range of strong convective cloud clusters; severity scores can be aggregated by regional mean or weighted (by the proportion of cold cloud cores) to serve as the final severity of the cloud cluster.

[0058] Through the embodiments of this application, while utilizing spatial neighborhood information to improve the continuity of discrimination, a physical constraint mechanism is used to ensure that information aggregation conforms to the consistency of the environmental background, making the probability and severity scores of strong convection more coordinated in space and more stable under different backgrounds, and can directly form an operational identification output of strong convective cloud clusters in terms of location range and severity classification.

[0059] Regarding the training of the physical constraint graph convolutional model, in some examples of embodiments of this application, a training sample set can be constructed based on historical operational samples. Each sample in the training sample set includes at least: a multi-spectral satellite remote sensing observation data sequence arranged in chronological order, environmental parameter data aligned with the observation data in time and space, a candidate cloud cluster graph structure and its node feature vectors constructed from a set of candidate cloud cluster regions, and supervision information corresponding to the candidate cloud cluster regions. The supervision information can be generated by manual annotation results, spatiotemporal matching results of multi-source observation products such as radar echoes / lightning location / ground disaster records, or historical identification results from operational systems, and is used to indicate the strong convection category label and / or severity label of the candidate cloud cluster regions. During the training process, the model parameters are optimized and updated based on the supervision information, and the classification error of the strong convection probability, the fitting error of the severity score, and the physical consistency constraint penalty term related to environmental physical features can be considered simultaneously in the loss term, so that the model can maintain the constraint logic of the graph convolutional neighborhood aggregation process modulated by environmental physical features while learning the multi-channel visual representation of candidate clouds. It should be noted that model training can be implemented in various ways, such as offline training, incremental updates, or transfer adaptation, depending on the actual engineering conditions, and this does not impose any restrictions on the training strategy.

[0060] Regarding the implementation details of generating multi-channel observation data sequences, in some examples of embodiments of this application, radiometric calibration is performed on multi-spectral satellite remote sensing observation data to convert the observation values ​​of each spectral band into brightness temperature or reflectance data.

[0061] In some implementations, for infrared and water vapor absorption channels, digital quantization values ​​or radiance can be corrected according to sensor calibration coefficients, and brightness temperature data can be obtained by Planck function inversion; for visible or near-infrared channels, the observed values ​​can be converted into top atmospheric reflectance or normalized reflectance, while quality control and mask marking are performed on bad pixels, strip noise, abnormal saturation and missing pixels to ensure that the data from different channels are comparable in terms of physical dimensions and numerical distribution.

[0062] Then, based on the atmospheric stratification information in the environmental parameter data and the satellite observation geometric parameters, parallax correction and georegistration are performed on the observation data of each spectral band, so that the same pixel at different times and in different spectral bands is mapped to a unified spatial grid.

[0063] In some implementations, atmospheric stratification information can be used to determine the representative height of cloud bodies (e.g., using tropopause height, temperature profiles, or external cloud top height estimates). Combined with the scanning angle, azimuth angle, and pixel geolocation model of satellite imaging, the parallax displacement of high-cloud-top targets is corrected, adjusting the pixel observation position from a "line-of-sight intersection" to a "surface-projection-consistent position." Subsequently, observations from different spectral bands (which may have different spatial resolutions and projection methods) are resampled to a unified spatial grid (e.g., a unified latitude and longitude grid or a unified geosynchronous projection grid). Continuous physical quantities can be interpolated using bilinear / bicubic interpolation, and mask-like or discrete markers can be mapped using nearest neighbor methods to achieve consistent mapping of corresponding pixels across spectral bands. This significantly reduces the spatial misalignment of high-cloud-top convective clouds in different channels and at different times, improves the consistency of key structures such as candidate cloud region outlines and the coldest cloud top regions across multiple channels, and thus reduces missegmentation and misjudgment caused by geometric errors.

[0064] Then, the registered spectral data were time-aligned according to the observation time sequence and organized into a multi-channel observation data sequence including the water vapor channel, infrared window channel and mid-infrared channel.

[0065] In some implementations, for channels with scanning time differences, time interpolation (linear / spline) can be performed on key channels to obtain multi-channel frames consistent with the target observation time; missing times can be marked and masked statistical processing can be performed in subsequent feature calculations. This ultimately forms a data structure arranged along the time dimension (e.g., a tensor of the form T×H×W×C), where the channel dimension can include water vapor channels, infrared window channels, and mid-infrared channels, etc., to simultaneously characterize cloud top thermal state, water vapor background differences, and microphysical / cloud facies information. This ensures stable temporal continuity and channel consistency of the model input and improves the identifiability of strong convection development stages.

[0066] Regarding the determination process of the adaptive dynamic threshold in the embodiments of this application, in some examples of the embodiments of this application, the tropopause temperature at each pixel coordinate is extracted from the environmental parameter data. and relative humidity factor The standard deviation of infrared brightness temperature of the multi-channel observation data sequence within a local window is calculated. .

[0067] here, The relative humidity profile in the environmental parameters can be averaged and normalized within a preset pressure layer (e.g., 700-300 hPa) according to weights, so that its value falls within a stable range of [0,1] or [0,100], so that it can be used as a gating variable in subsequent threshold adjustment; at the same time, the infrared window brightness temperature (which can be the infrared window channel brightness temperature at the target observation time, or a set of brightness temperatures at multiple times within a short time window) for characterizing the cloud top thermal condition is selected from the multi-channel observation data sequence. A preset neighborhood window centered on the center (e.g.) Calculate the standard deviation of infrared brightness temperature within a pixel. This standard deviation is used to characterize the spatial (or spatiotemporal) fluctuations in local cloud top brightness temperature: when cold cloud cores, anvil-shaped cloud boundaries, and convective cells exist side-by-side within the cloud, the brightness temperature within the window exhibits stronger inhomogeneity, thus... Increase.

[0068] Then, an environment-adaptive thresholding function is constructed to calculate the adaptive dynamic threshold for each pixel. :

[0069] Equation (1)

[0070] In the formula, Represents pixel coordinates, and The preset empirical adjustment coefficient, For The standard deviation of infrared brightness temperature within the preset neighborhood window centered on the center. pixel coordinates The relative humidity factor at a given location. Here, the adaptive dynamic threshold decreases as the relative humidity factor increases, in order to improve the sensitivity of deep convective clouds in high humidity environments.

[0071] in, and Preset empirical adjustment coefficients are used to control the influence of "local brightness temperature fluctuation" and "humidity background" on the threshold, respectively; and Having the same temperature dimensions, A temperature correction term is formed for the threshold based on the cloud structure complexity; When it is a dimensionless or normalized quantity, Used to map the humidity background to an equivalent temperature correction magnitude.

[0072] In equation (1), with As a reference for the upper limit of cooling that cloud tops may reach, the threshold is naturally shifted with changes in region, season, and tropopause; at the same time, the fluctuations in local cloud top brightness temperature are more pronounced. When it is larger, by subtracting This lowers the threshold, thereby improving the response to convective clouds with cold cores and strong structural gradients; when the ambient humidity background is higher. When it is larger, by subtracting The threshold is further lowered so that the algorithm has higher recognition sensitivity in high humidity and convection environments where it is easier to develop and maintain, while the threshold is relatively raised in relatively dry or unfavorable environments to suppress unnecessary candidate expansion.

[0073] Here, the obtained pixel-level dynamic threshold field can be directly used for subsequent cloud region segmentation / candidate extraction (e.g., determining that the brightness temperature of the infrared window region is lower than a certain value). (Pixels are used as candidate cloud regions and connected component aggregation is performed). Therefore, the threshold adopts an update mechanism that is adaptive to the environmental background, and is also adaptively adjusted according to the tropopause constraint, local cloud top structure fluctuations and humidity background, so as to maintain a more stable segmentation scale under different climatic backgrounds and capture deep convective cold cloud top targets more sensitively in high humidity environments.

[0074] Regarding the process of determining the candidate cloud region set, in some examples of embodiments of this application, an initial binary mask is generated based on an adaptive dynamic threshold, and infrared brightness temperatures lower than the adaptive dynamic threshold are selected. The area is marked as the initial selection area.

[0075] In the process of determining the candidate cloud cluster region set, in some examples of the embodiments of this application, the brightness temperature field of the infrared window channel at the time of target observation is first used as the basis. With pixel-level adaptive dynamic threshold Generate initial binary mask For example, the initial selected pixels can be labeled according to the following rules:

[0076] Equation (2)

[0077] In the formula, This is an indicator function.

[0078] More specifically, missing / low-quality pixels can be directly invalidated and removed from the threshold discrimination; when the spatial resolution of satellite channels is inconsistent, brightness temperature data already registered to a unified grid is preferred for comparison to ensure pixel-by-pixel alignment between the threshold field and the observation field. Therefore, "sufficiently cold cloud tops" are used as the candidate recall entry point, and through... By binding the threshold to the local environmental state, the initial mask maintains a consistent screening scale under different backgrounds, providing a high-coverage candidate set for subsequent fine screening.

[0079] Then, the brightness temperature difference between the water vapor channel and the infrared window channel, as well as the brightness temperature difference between the mid-infrared channel and the infrared window channel, are used to perform differential screening on the initially selected area in order to eliminate low-temperature cloud pixels with non-convective properties.

[0080] Specifically, the brightness temperature difference between the water vapor channel and the infrared window channel is calculated. And the brightness temperature difference between the mid-infrared channel and the infrared window channel. and in Differential constraints are applied to pixels with a value of 1 to form a filtered mask. .

[0081] Difference constraints can be implemented in two ways: empirical intervals or adaptive intervals. One way is to set... and One approach is to select candidates within a range of values ​​(e.g., only those falling within a preset upper and lower limit range are retained); another approach is to dynamically adjust the aforementioned range based on ambient humidity and tropopause conditions, so that the differential discrimination also shifts with changes in the background. Therefore, by utilizing the difference in water vapor absorption and mid-infrared sensitivity to cloud top height, cloud phase / particle size, and upper-layer humidity, the candidate range is further narrowed, reducing the probability of non-convective cold clouds (such as some stratiform clouds and widespread cold cirrus clouds) being included in the initial selection stage, thereby improving the "convective purity" of the candidates.

[0082] Then, morphological closing operations are performed on the filtered mask using structuring elements to fill the voids inside the cloud and connect adjacent broken patches.

[0083] More specifically, regarding masks Perform a morphological closing operation to obtain a more structurally complete cloud object. Specifically, select a structuring element that matches the data resolution. (For example, radius is) The disk structural element or side length is (a square structural element), which first expands and then erodes the mask. , This is a morphological dilation operator. This is a morphological erosion operator. Dilation connects adjacent patches separated by noise or local threshold fluctuations, while erosion restores the target boundary and suppresses excessive dilation after connection. The structuring element scale can be set according to the satellite pixel size, ensuring it only fills small holes within the cloud cluster and bridges nearby breaks, without excessively merging clearly separated cloud clusters. This improves the coherence and topological integrity of the candidate cloud cluster mask, provides a more stable target boundary for subsequent connected component segmentation, and reduces candidate number inflation caused by fragmentation.

[0084] Then, connected component analysis is performed on the morphologically processed mask to calculate the geometric area of ​​each connected component, and connected components with an area greater than a preset minimum area threshold are retained as candidate cloud regions.

[0085] Specifically, the morphologically processed mask Connectivity analysis is performed to form a set of candidate cloud regions. In implementation, 8-neighbor connectivity markers can be used to obtain each connected component. Given a set of pixels, calculate its geometric area. (in The number of connected cells. (This represents the actual area corresponding to a single pixel, which can be calculated from the projection and resolution); then, area threshold filtering is performed to retain pixels that meet the requirements. The connected components are used as candidate cloud regions, and object-level attributes such as the bounding rectangle, centroid coordinates, and boundary contours of each region can be output simultaneously.

[0086] Thus, the pixel-level mask is transformed into a computable and traceable "object-level" candidate cloud unit, filtering out small-scale noise targets that are not of business significance, and finally obtaining a set of candidate cloud regions with clear structure, controlled scale and easy subsequent graph model processing.

[0087] Regarding the process of generating node feature vectors for candidate cloud regions, in some examples of embodiments of this application, the standard deviation of infrared channel brightness temperature within the candidate cloud region is calculated. Brightness difference between water vapor and infrared channel The box-dimensional algorithm is used to estimate the fractal dimension that characterizes the surface roughness of cloud tops. And using the optical flow algorithm to calculate the area growth rate characterizing the temporal evolution. .

[0088] In some implementations, for each candidate cloud region obtained through connected component analysis (It consists of a set of spatially connected pixels), and in multi-channel observation data on a unified grid, pixel masking statistics are performed on pixels within a region to extract image visual features. Specifically, this can be done in the brightness temperature field of the infrared window channel. Standard deviation of brightness temperature in the calculation region ,For example This is used to characterize the brightness temperature fluctuations and internal structural inhomogeneities of cloud tops; among them, This is a standard deviation calculation unit. It also calculates the brightness temperature difference between the water vapor and infrared channels. ,For example Alternatively, quantile differences can be used to improve sensitivity to cold cloud nuclei.

[0089] To characterize the geometric features of the cloud top surface's "roughness / complex boundary," the fractal dimension can be estimated using a box-counting algorithm on the infrared brightness temperature field within the region or on its edge / isotherm binary map. Projecting the region onto a binary or multi-threshold edge map (e.g., for...) Gradient or edge detection yields a set of boundary pixels, within a multi-scale grid size. The number of boxes required to cover the target set is calculated below. And through linear fitting and The slope is obtained This is to reflect the multi-scale complexity of cloud boundaries and internal textures.

[0090] Furthermore, to reflect the intensity of temporal evolution, optical flow estimation (such as dense optical flow) can be performed on the region based on observation frames from adjacent time points. Forward mapping of optical flow field to time. The predicted region was then obtained. And use this to calculate the area growth rate. Alternatively, the expansion trend can be characterized by the average value of the optical flow outward expansion component at the boundary. The above feature calculations can all be completed at the regional scale, and normalization processing can be introduced for different resolutions / different regional scales to ensure that the features are comparable among different samples. Thus, the observable structure and evolution state of candidate clouds can be characterized from four complementary dimensions: thermal fluctuations, absorption differences, geometric complexity, and development rate.

[0091] Then, the convective available potential energy corresponding to the candidate cloud cluster region space is mapped from the environmental parameter data. and convection suppression energy It can be extracted by interpolation based on the centroid location of the region, or the mean / quantile aggregation of environmental parameters within the region coverage area can be performed to improve robustness.

[0092] Furthermore, a nonlinear fusion function is employed to fuse image visual features with environmental physical features in order to construct a convection complexity index:

[0093] Equation (3)

[0094] In the formula, As a measure of convection complexity, This is a preset positive lower bound constant used to impose a lower bound constraint on the denominator; to These are the weight coefficients for the corresponding feature terms.

[0095] In equation (3), This is used to retain only the positive water vapor difference contribution, which is more indicative of deep convection, and to avoid the negative difference causing undesirable offsetting effects on the indicators; Expressing the contrast between "potentially releaseable energy" and "inhibition strength", Used to impose a lower bound constraint on the denominator to avoid... The values ​​become unstable when the values ​​are very small or missing. to The weighting coefficients, which can be set empirically or learned through data-driven methods during the training phase, are used to adjust the contribution of different components to the final complexity. Finally, these five variables are concatenated (and auxiliary quantities such as region area and coldest brightness temperature can be added) to form the node feature vector of the candidate cloud region. Thus, through the nonlinear coupling of "visual structure and temporal evolution" and "environmental supportability," the node representation simultaneously reflects the complex development characteristics of the cloud itself and the background thermal constraints.

[0096] Regarding the calculation of fractal dimension and area growth rate In some examples of embodiments of this application, the box dimension estimation method is used to calculate the fractal dimension. By changing the grid scale For a binarized image covering the candidate cloud region, count the number of grid cells containing cloud pixels. .

[0097] In some implementations, after the candidate cloud region extraction is completed, the target observation time can be used as the starting point. Binarized images for fractal analysis are generated from the candidate cloud region mask. (For example, the pixel value in the candidate area is set to 1 and the background value is set to 0; alternatively, morphological refinement or edge extraction can be performed on the cloud boundary before statistical analysis to better highlight the irregularity of the boundary).

[0098] Subsequently, box-counting was used to estimate the fractal dimension. Specifically, a set of discrete grid scales was selected. (For example, taking in units of pixels) Pixel, (and limit the maximum scale to no more than the shorter side length of the bounding rectangle of the candidate region), at each scale The binary image is divided into sections with side lengths of... Count the number of non-empty grid cells that "contain at least one cloud pixel (value 1)". Expressed in limit form:

[0099] Equation (4)

[0100] In the formula, The grid's side length scale. The number of non-empty grids required to cover the cloud region at this scale; fractal dimension. It is used to quantitatively characterize the irregularity of cloud edges and the complexity of the filling space.

[0101] In equation (4), the discrete image can be approximated by linear fitting of a multi-scale point array. For the vertical axis, Using the horizontal axis as the x-axis, the slope is obtained by least-squares fitting. The estimated value; of which The size increases with decreasing scale. If the cloud edges are more intricate and the internal filling is more complex, then more non-empty grid coverage is required at smaller scales, thus... Slower decay and steeper slope indicate higher spatial complexity. Therefore, the "degree of irregularity of cloud edges and complexity of spatial filling" can be quantified into a scale-independent index, enabling subsequent models to distinguish cloud objects with more complex structures and greater convective organization characteristics.

[0102] Then, the area growth rate is calculated based on the changes in cloud area at adjacent observation times. :

[0103] Equation (5)

[0104] In the formula, The pixel area of ​​the candidate cloud region at the current observation time. This represents the pixel area of ​​the cloud cluster corresponding to the candidate cloud cluster region at the previous observation time. The time interval between adjacent observation times; when The time indicates that the cloud cluster is in the stage of enhanced convection development.

[0105] In equation (5), the area growth rate The calculation is based on the changes in candidate cloud area between adjacent observation times. Specifically, it can be set as follows: For a moment The pixel area of ​​the candidate cloud region (number of pixels multiplied by the actual area of ​​a single pixel, or directly using the number of pixels as the relative area), and at the previous observation time. The area corresponding to the candidate cloud cluster in space was determined. This correspondence can be achieved through regional centroid nearest neighbor matching, regional overlap ratio (IoU) maximum matching, or by combining environmental wind field / optical flow for short-term advection extrapolation and then matching, in order to reduce mismatches caused by cloud movement.

[0106] For example, For the time interval between two observations (e.g., 10 minutes, 15 minutes), the denominator of equation (5) contains... Used to normalize area changes into relative growth rates Used to obtain the growth rate per unit time; when This indicates that the candidate cloud clusters show an expanding trend within this time window, which typically corresponds to the intensification phase of convective development. This corresponds to the maintenance or weakening / dissipation phase.

[0107] Thus, the "temporal evolution intensity" of cloud clusters can be explicitly characterized in a comparable rate, enabling subsequent identification to not only rely on single-moment morphology but also utilize short-term evolution information to enhance sensitivity to the development process of strong convection.

[0108] Regarding the implementation details of constraining neighborhood feature aggregation, in some examples of embodiments of this application, an energy attention mechanism is introduced to dynamically adjust the feature aggregation weights of neighboring nodes during the information transmission process of the graph convolutional layer.

[0109] Here, in the process of graph convolution information transmission in the candidate cloud map, in order to make the propagation of neighborhood information conform to the constraint logic of atmospheric thermodynamic background consistency, an "energy-attention" mechanism is introduced before the traditional neighborhood aggregation to dynamically weight the contribution of each neighbor node.

[0110] Then, compute the nodes. Its neighboring nodes Physical energy weighting coefficients between :

[0111] Equation (6)

[0112] In the formula, Represents a node The set of neighboring nodes, and Representing nodes respectively and Convective available potential energy, and Representing nodes respectively and Convection suppression energy, These are the energy modulation parameters.

[0113] In equation (6), for graph nodes Convective potential energy is obtained in advance by mapping its environmental physical characteristics. With convection suppression energy (For example, values ​​can be interpolated at corresponding node locations from numerical weather prediction / reanalysis fields, and missing values ​​can be masked or replaced with regional statistics), and with As a physical energy index characterizing the "relative suppression strength of energy support for triggerable convection", among which It is a positive lower bound constant, used to avoid The small value leads to numerical instability. Subsequently, a normalized exponential form (softmax) is used to evaluate the nodes. neighborhood set Calculate the physical energy weighting coefficient for each neighbor. .

[0114] here, Used to control the "sharpness" of the weight distribution. The larger the value, the more the model tends to highlight neighbors with higher energy indicators; The smaller the value, the smoother the weighting and the closer it is to average aggregation. Due to the above... Normalized to 1 within the neighborhood, it can be directly used as the physical gating coefficient for neighborhood aggregation, allowing neighboring nodes in more favorable convective thermal backgrounds (high CAPE, low CIN) to have a higher proportion in information transmission, while the contribution of neighbors with insufficient energy support or significant inhibition to aggregation is suppressed, thereby reducing the propagation of inconsistent information across unfavorable backgrounds from a mechanistic perspective.

[0115] Furthermore, nodes are updated based on physical energy weighting coefficients. In the Hidden state of layers :

[0116] Equation (7)

[0117] In the formula, For network layer indexing, and The first Layer weight matrix and bias, For activation function, For nodes In the The hidden state vector of the layer.

[0118] In equation (7), The ReLU activation function can be used. Alternatively, a self-loop term can be added (to... (It considers itself as one of its neighbors) to maintain node self-information, and can... Truncation / temperature scaling is applied to enhance numerical stability.

[0119] Through the above-mentioned coupling of "physical energy weights - weighted aggregation", the graph convolutional layer can prioritize the aggregation of neighborhood information that is more supportive of strong convection under thermal conditions while maintaining the utilization of spatial proximity information. This makes the strong convection probability and severity score output by the model more consistent with physical conditions, more robust to environmental field abrupt changes, and improves the inference stability and adaptability in different regions and under different climatic backgrounds.

[0120] In some examples of embodiments of this application, after outputting the strong convective cloud cluster identification result, the method further includes spatiotemporal consistency constraint processing. To ensure that the strong convective cloud cluster identification result meets the requirements of operational continuity and evolutionary consistency, at the target observation time output by the model... After identifying the severe convection, the first candidate cloud cluster region was also identified as having severe convection. Perform spatiotemporal consistency constraint processing.

[0121] Specifically, firstly, the first candidate cloud cluster region identified as a strong convective cloud cluster at the target observation time is selected. And construct a time window that includes the target observation time.

[0122] Then, the set of candidate cloud cluster regions and their corresponding strong convection probabilities for each adjacent observation time within the time window are obtained, and a search is performed on the first candidate cloud cluster region. Matching candidate cloud regions .

[0123] Here, around Construct a time window (e.g., extend it forward / backward by several observation periods), and then perform adjacent observations within the window. Obtain the set of candidate cloud cluster regions and their probability of strong convection. Then, use... Starting from the point of reference, search for matching regions in adjacent time intervals. The matching process calculates the spatiotemporal matching cost. Implementation. For example, calculate according to the following formula. and Spatiotemporal matching cost between :

[0124] Equation (8)

[0125] In the formula, for and Spatial overlap between them and They are respectively and The position of the center of mass, for The velocity vector of motion, This represents the time difference between adjacent observation times and the target observation time. For normalized distance factor, for and The angle between the directions of motion.

[0126] In equation (8), the first term uses spatial overlap (IOU) to measure the consistency of the region's morphology; the greater the overlap, the lower the cost. The second term uses the consistency of the centroid position to constrain the continuity of the target's motion. and These are the regional centroids, For candidate regions in The velocity vector at any given moment (which can be estimated from the centroid displacement at adjacent moments, or from cloud advection by optical flow / wind field). Due to time difference, This represents the predicted centroid obtained by extrapolation based on motion. To normalize the distance factor and eliminate the dimensional effects of different resolutions / region scales; the third term uses the angle between the motion directions. Suppressing illogical matching caused by directional mutations The closer it is to 1, the more consistent the direction and the lower the cost. The matching weights are used to balance the three types of constraints: form, displacement, and direction. When the spatiotemporal matching cost is less than a preset matching threshold, it is considered... and They belong to the same evolutionary cloud cluster in space and time. join in The corresponding cross-time association chain can be selected from multiple candidates at the same time according to the principle of "lowest cost priority / maximum overlap priority" to avoid forks.

[0127] Next, the duration of the cross-time association chain within the time window is calculated. (For example, the difference between the first and last timestamps of the chain or the number of observation frames covered by the chain multiplied by the observation period), and calculate the cumulative confidence score based on the strong convection probability of each node on the chain. .

[0128] Equation (9)

[0129] In the formula, For the index of a node in a time-series association chain, The chain length of the chain that is associated across time points. The timestamp corresponding to the time of the target observation. For the first time-related chain The observation timestamps corresponding to each node For the first time-related chain The probability of strong convection at each node This is the time decay coefficient.

[0130] In equation (9), This is a time decay term, used to emphasize evidence closer to the target time and weaken uncertain contributions from times farther away from the target time. This is the time decay coefficient used to control the decay rate.

[0131] Furthermore, if the duration Reaching a preset duration threshold and / or cumulative confidence level If the preset threshold is reached, the strong convective cloud cluster identification result of the first candidate cloud cluster region is confirmed as valid, and a continuous early warning signal containing spatiotemporal evolution information is output; otherwise, the strong convective cloud cluster identification result of the first candidate cloud cluster region is marked as instantaneous noise.

[0132] Finally, if Reaching a preset duration threshold, and / or If a preset threshold is reached, the strong convection identification result for the first candidate cloud cluster region is confirmed to have continuous support in the time dimension, and a continuous early warning signal containing spatiotemporal evolution information is output (e.g., duration, trajectory direction / velocity, probability changes over time, etc.). Otherwise, the result is marked as instantaneous noise and the output is suppressed. Thus, without changing the front-end identification recall, the temporal stability and evolutionary consistency of the results can be significantly enhanced, reducing instantaneous alarms caused by single-frame anomalies, short-term false detections, or fragmented segmentation, thereby improving the operational availability of strong convection monitoring products.

[0133] To visually present the results of the strong convective cloud cluster identification, Figure 2 A schematic diagram of a true-color composite image of the area to be monitored at the target observation time is shown; Figure 3 This shows the true-color composite image (i.e. Figure 2 A schematic diagram showing the strong convective cloud cluster identification results superimposed on the method of this application; Figure 4 It shows the relationship with Figure 2 A schematic diagram of the long-wave infrared brightness temperature map of the same region at the same time; Figure 5 The image shows the brightness temperature map in the long-wave infrared region (i.e. Figure 4 The diagram shows the strong convective cloud cluster identification results (outlines / boundaries) superimposed on different base maps. By superimposing cloud cluster identification labels (outlines / boundaries) on different base maps, the spatial range of the identified strong convective cloud clusters (or candidate cloud cluster regions) can be intuitively verified from two perspectives: visible light appearance and infrared cloud top thermal characteristics.

[0134] Furthermore, Figure 3 and Figure 5 The outline of the superimposed annotation corresponds to the boundary of the cloud object determined by this application after candidate extraction and classification reasoning: it is represented by a cloud area morphology consistent with thick convective cloud body on the true color image, and has a good spatial correspondence with the low brightness temperature (cold cloud top) region of the cloud top on the long-wave infrared image, thereby verifying the consistency between the geographical range of the strong convective cloud cluster output by this application and the cooling characteristics of the cloud top, and providing a visualization basis for subsequent severity scoring and spatiotemporal evolution tracking.

[0135] To verify the effectiveness of the proposed method, this paper will further elaborate on the details of the designed comparative experiment. Specifically, multi-spectral data from the FY-4B satellite (including at least the water vapor channel and infrared window channel, and mid-infrared / split window channel if necessary) of a target area from June to September 2023 were selected. The data were calibrated, uniformly projected, and spatiotemporally registered to construct a sample.

[0136] Supervised labels were generated using CloudSat-CALIPSO in conjunction with cloud classification products. These labels were then spatiotemporally matched with FY-4B observations according to preset time windows and spatial neighborhood criteria (e.g., limiting time difference and spatial distance thresholds and resampling to a unified grid), thus obtaining ground truth values ​​usable for pixel-level / region-level evaluation. Approximately 10,000 samples were constructed, with about 8,000 used for training and 2,000 for testing. To avoid leakage of the same convection process between training and testing, it was preferable to divide the samples by date / process rather than randomly dividing them frame by frame.

[0137] The baselines for comparison included Infrared Window Thresholding (IRW), Brightness-Temperature Difference (BTD), Tropopause Differential (TROPO), and lightweight network ensemble algorithms (such as ENet and ESPNet ensemble models). All baselines were run on the same test set and under consistent input channel / preprocessing conditions.

[0138] Evaluation metrics include accuracy (ACC), recall, false alarm rate (FAR), and Kappa coefficient. To address class imbalance, stratified sampling / class weighting methods can be used for training or evaluation correction. Bootstrap resampling is employed to repeatedly sample test set metrics to estimate means and confidence intervals, thereby improving statistical robustness. Furthermore, ablation experiments are conducted, removing the CCI metric, graph structure modeling, or physical attention constraint module, respectively, to construct control models and quantify the contribution of each key module to performance improvement.

[0139] Figure 6 The comparison results of detection accuracy of different methods with varying brightness temperature thresholds are shown. The horizontal axis represents the infrared brightness temperature threshold (K, ranging from approximately 200–240 K), and the vertical axis represents accuracy. The curves correspond to the Infrared Window Thresholding (IRW), Brightness Temperature Difference (BTD), Tropopause Differential (TROPO), and the Physics-Constrained Graph Convolutional Network + Convective Complexity Index (PC-GCN+CCI) method presented in this paper.

[0140] like Figure 6As shown, the accuracy of the three traditional thresholding methods—IRW, BTD, and TROPO—fluctuates significantly with changes in the threshold and each has its own "optimal threshold" range. For example, IRW reaches a peak around 215 K and then rapidly decreases as the threshold increases. BTD and TROPO also peak around 220–225 K, and their accuracy decreases significantly when the threshold deviates from this range. The overall highest accuracy does not exceed 0.80 (approximately 0.70–0.72 in the figure), indicating that their performance is quite sensitive to threshold settings. In contrast, the PC-GCN+CCI curve remains around 0.88–0.90 throughout the entire threshold scanning range, with very little variation with the threshold, demonstrating stronger stability and robustness. This result reflects, experimentally, that the proposed scheme reduces the dependence on a single fixed brightness temperature threshold through dynamic thresholding and physical constraint mechanisms, thus making the recognition accuracy insensitive to threshold perturbations.

[0141] Figure 7 The diagram shows a heatmap of the convective complexity index (CCI) distribution over time for several candidate cloud clusters selected from the experimental test set. The horizontal axis represents the time step, the vertical axis represents the candidate cloud cluster number, and the color bars indicate the magnitude of the CCI: the color changes from dark (purple / blue) to light (green / yellow) corresponding to a CCI from low to high.

[0142] like Figure 7 As shown, some candidate cloud clusters maintain high CCI (brighter bands) across multiple consecutive time steps, exhibiting strong and persistent convective structure characteristics; while other candidate cloud clusters show persistently low CCI or only short-term increases (darker or scattered bright bands), reflecting weaker convective complexity or discontinuous evolution. This result demonstrates that CCI can effectively characterize the differences in convective strength and persistence among different candidate cloud clusters in both spatiotemporal dimensions, providing stable and interpretable feature support for subsequent graph-based discrimination.

[0143] Furthermore, to further verify the effectiveness of the method, this paper also statistically analyzed the overall evaluation metrics (such as ACC, Recall, FAR, and Kappa) of each algorithm on the test set. Compared with the baseline algorithm, the proposed method outperforms the baseline algorithm in terms of accuracy, recall, and Kappa coefficient, and has a lower false alarm rate, thus demonstrating the effectiveness of introducing physical features and graph convolutional inference mechanisms in the identification of strong convective clouds from an overall performance perspective.

[0144] This paper addresses the need for satellite remote sensing identification of strong convective clouds by proposing a physics-guided method. This method uses an adaptive dynamic threshold driven by environmental parameters to extract candidate cloud regions, and combines differential filtering, morphological processing, and connected component analysis to obtain candidate cloud objects. At the object level, it integrates visual features from multi-spectral observations with environmental physical features to construct a Convective Complexity Index (CCI) to characterize convective structure and development intensity. Furthermore, it constructs a graph structure based on the spatial adjacency relationships of candidate clouds, and employs a physically constrained graph convolutional network to introduce energy / environment consistency constraints during inference, achieving joint discrimination and output of the probability and severity of strong convection in candidate clouds, thus balancing identification accuracy, stability, and physical interpretability.

[0145] In experimental verification, under the dataset and evaluation index system used, this method shows better overall performance (such as improved accuracy, recall, and Kappa, and reduced false alarm rate) compared with traditional thresholding methods and the comparison deep learning baseline. At the same time, it demonstrates better robustness and generalization ability under threshold perturbation and background changes.

[0146] In the future, multi-source observations such as satellite lightning imagers can be further integrated to enhance the characterization of convection intensity, and coupling with numerical weather prediction models can be explored to support short-term forecasts of the occurrence and development of severe convection. At the same time, engineering deployment and real-time evaluation on operational platforms can be promoted to test their online operating efficiency, stability and long-term applicability.

[0147] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of combined actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Secondly, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application. In the above embodiments, the descriptions of each embodiment have their own emphasis; for parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0148] Figure 8 A structural block diagram of an example of a strong convective cloud cluster identification system based on satellite remote sensing data according to an embodiment of this application is shown.

[0149] like Figure 8 As shown, the strong convective cloud cluster identification system 800 based on satellite remote sensing data includes a data acquisition unit 810, an adaptive candidate extraction unit 820, a physical guided feature construction unit 830, a graph structure construction unit 840, and a physical constraint reasoning unit 850.

[0150] The data acquisition unit 810 is used to acquire multi-band satellite remote sensing observation data of the area to be monitored within at least one observation period, and to acquire environmental parameter data corresponding to the multi-band satellite remote sensing observation data in time and space; and to perform spatiotemporal registration preprocessing on the multi-band satellite remote sensing observation data to generate a multi-channel observation data sequence arranged in chronological order.

[0151] The adaptive candidate extraction unit 820 is used to determine an adaptive dynamic threshold for cloud segmentation based on the environmental parameter data, and to perform candidate extraction processing on the cloud area at the target observation time in the multi-channel observation data sequence using the adaptive dynamic threshold to obtain a set of candidate cloud areas, wherein each candidate cloud area in the set of candidate cloud areas corresponds to a set of spatially connected pixel areas.

[0152] The physical guidance feature construction unit 830 is used to extract image visual features based on the multi-channel observation data sequence for each candidate cloud region in the candidate cloud region set, and determine environmental physical features based on the environmental parameter data. Based on the image visual features and the environmental physical features, it constructs a convection complexity index to characterize the degree of strong convection in the candidate cloud region, and generates a node feature vector for the candidate cloud region. The node feature vector includes the image visual features, the environmental physical features, and the convection complexity index.

[0153] The graph structure construction unit 840 is used to construct a candidate cloud cluster graph structure based on the set of candidate cloud cluster regions. The graph nodes of the candidate cloud cluster graph structure correspond to each candidate cloud cluster region, and the graph edges are established between candidate cloud cluster regions that meet the preset distance conditions, which are used to characterize the spatial proximity relationship of the candidate cloud cluster regions.

[0154] The physical constraint inference unit 850 is used to input the candidate cloud cluster graph structure and the node feature vectors of each graph node into the physical constraint graph convolutional inference model to output the strong convection probability and severity score corresponding to each candidate cloud cluster region, and output the strong convection cloud cluster identification result based on the strong convection probability and the severity score. The strong convection cloud cluster identification result includes at least the geographical range of the strong convection cloud cluster and the corresponding severity score. In the graph convolutional inference process, the physical constraint graph convolutional inference model constrains the aggregation of neighborhood features based on the environmental physical features.

[0155] In some embodiments, this application provides a non-volatile computer-readable storage medium storing one or more programs including execution instructions. The execution instructions can be read and executed by electronic devices (including but not limited to computers, servers, or network devices) to perform the steps of any of the above-described methods for identifying strong convective cloud clusters based on satellite remote sensing data.

[0156] In some embodiments, this application also provides a computer program product, the computer program product including a computer program stored on a non-volatile computer-readable storage medium, the computer program including program instructions, which, when executed by a computer, cause the computer to perform the steps of any of the above-described methods for identifying strong convective cloud clusters based on satellite remote sensing data.

[0157] In some embodiments, this application also provides an electronic device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of a method for identifying strong convective cloud clusters based on satellite remote sensing data.

[0158] The above-described product can perform the methods provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects for performing the methods. Technical details not described in detail in this embodiment can be found in the methods provided in the embodiments of this application.

[0159] The electronic devices in this application can exist in various forms, including but not limited to: mobile communication devices, ultra-mobile personal computer devices, portable entertainment devices, or other airborne electronic devices with data interaction functions.

[0160] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0161] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0162] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for identifying strong convective cloud clusters based on satellite remote sensing data, characterized in that, The method includes: Acquire multispectral satellite remote sensing data of the area to be monitored within at least one observation period, and acquire environmental parameter data corresponding to the multispectral satellite remote sensing data in time and space; perform spatiotemporal registration preprocessing on the multispectral satellite remote sensing data to generate a multi-channel observation data sequence arranged in chronological order; the multispectral satellite remote sensing data includes visible light channel data, infrared window area channel data, and water vapor channel data; the environmental parameter data includes at least one of the following: temperature and humidity profile parameters characterizing the convective environment, convective effective potential energy parameters, lifting index parameters, tropopause-related parameters, and vertical wind shear parameters; The adaptive dynamic threshold for cloud segmentation is determined based on the aforementioned environmental parameter data, specifically including: The tropopause temperature at each pixel coordinate is extracted from the environmental parameter data. and relative humidity factor And calculate the standard deviation of infrared brightness temperature of the multi-channel observation data sequence within a local window. ; Construct an environment-adaptive threshold function to calculate the adaptive dynamic threshold for each pixel. : ， In the formula, Represents pixel coordinates, and The preset empirical adjustment coefficient, For The standard deviation of infrared brightness temperature within the preset neighborhood window centered on the center. pixel coordinates The relative humidity factor at that location; The adaptive dynamic threshold is used to perform candidate extraction processing on the cloud area at the target observation time in the multi-channel observation data sequence to obtain a set of candidate cloud cluster regions, wherein each candidate cloud cluster region in the set of candidate cloud cluster regions corresponds to a set of spatially connected pixel regions. For each candidate cloud region in the candidate cloud region set, visual features of the images are extracted based on the multi-channel observation data sequence, and environmental physical features are determined based on the environmental parameter data. A convection complexity index is constructed based on the visual features of the images and the environmental physical features to characterize the degree of strong convection in the candidate cloud region, specifically including: Calculate the standard deviation of infrared channel brightness temperature within the candidate cloud region. Brightness difference between water vapor and infrared channel The box-dimensional algorithm is used to estimate the fractal dimension that characterizes the surface roughness of cloud tops. And using the optical flow algorithm to calculate the area growth rate characterizing the temporal evolution. ; Mapping the convective available potential energy corresponding to the regional space of the candidate cloud cluster from the environmental parameter data. and convection suppression energy ; A nonlinear fusion function is used to fuse the image visual features with the environmental physical features to construct a convection complexity index. ， In the formula, As a measure of convection complexity, This is a preset positive lower bound constant used to impose a lower bound constraint on the denominator; to These are the weight coefficients for the corresponding feature terms; Generate node feature vectors for the candidate cloud region; the node feature vectors include the image visual features, the environmental physical features, and the convection complexity index. A candidate cloud cluster graph structure is constructed based on the set of candidate cloud cluster regions. The graph nodes of the candidate cloud cluster graph structure correspond to each candidate cloud cluster region, and the graph edges are established between candidate cloud cluster regions that meet the preset distance conditions, which are used to characterize the spatial proximity relationship of candidate cloud cluster regions. The candidate cloud cluster graph structure and the node feature vectors of each graph node are input into the physical constraint graph volume integral class model to output the strong convection probability and severity score corresponding to each candidate cloud cluster region, and output the strong convection cloud cluster identification result based on the strong convection probability and the severity score. The strong convection cloud cluster identification result includes at least the geographical range of the strong convection cloud cluster and the corresponding severity score. Specifically, the physical constraint graph convolutional inference model constrains the aggregation of neighborhood features based on the environmental physical features during the graph convolutional inference process, including: During the information transmission process of the graph convolutional layer, an energy attention mechanism is introduced to dynamically adjust the feature aggregation weights of neighboring nodes; compute nodes Its neighboring nodes Physical energy weighting coefficients between : ， In the formula, Represents a node The set of neighboring nodes, and Representing nodes respectively and Convective available potential energy, and Representing nodes respectively and Convection suppression energy, These are energy modulation parameters; Update nodes based on physical energy weighting coefficients In the Hidden state of a layer : ， In the formula, For network layer indexing, and The first Layer weight matrix and bias, For activation function, For nodes In the The hidden state vector of the layer.

2. The method according to claim 1, characterized in that, The process of performing spatiotemporal registration preprocessing on the multi-spectral satellite remote sensing observation data to generate a multi-channel observation data sequence arranged in chronological order includes: Radiometric calibration is performed on the multi-band satellite remote sensing data to convert the observations of each spectral band into brightness temperature or reflectance data; Based on the atmospheric stratification information and satellite observation geometric parameters in the environmental parameter data, parallax correction and georegistration are performed on the observation data of each spectral band, so that the same pixel at different times and in different spectral bands is mapped to a unified spatial grid. The registered spectral data were time-aligned according to the order of observation time and organized into a multi-channel observation data sequence including the water vapor channel, infrared window channel, and mid-infrared channel.

3. The method according to claim 1, characterized in that, The adaptive dynamic threshold is used to perform candidate extraction processing on the cloud region at the target observation time in the multi-channel observation data sequence to obtain a set of candidate cloud region areas, including: An initial binary mask is generated based on the adaptive dynamic threshold, and the infrared brightness temperature is lower than the adaptive dynamic threshold. The area is marked as the initial selection area; The initial selection area is differentially screened using the brightness temperature difference between the water vapor channel and the infrared window channel, as well as the brightness temperature difference between the mid-infrared channel and the infrared window channel, in order to eliminate low-temperature cloud pixels with non-convective properties. Morphological closing operations are performed on the screened mask using structuring elements to fill the voids inside the cloud and connect adjacent broken patches. Connectivity analysis is performed on the morphologically processed mask to calculate the geometric area of ​​each connected region, and connected regions with an area greater than a preset minimum area threshold are retained as candidate cloud regions.

4. The method according to claim 1, characterized in that, The fractal dimension and the area growth rate The calculation process specifically includes: The fractal dimension was calculated using the box dimension estimation method. By changing the grid scale The number of grid cells containing cloud pixels is counted from the binarized image covering the candidate cloud region. : ， In the formula, The grid's side length scale. The number of non-empty grids required to cover the cloud region at this scale; the fractal dimension. Used to quantitatively characterize the irregularity of cloud edges and the complexity of the filling space; The area growth rate is calculated based on the cloud area changes at adjacent observation times. : ， In the formula, The pixel area of ​​the candidate cloud region at the current observation time. This represents the pixel area of ​​the candidate cloud cluster corresponding to the spatial region of the previous observation time. The time interval between adjacent observation times; when The time indicates that the cloud cluster is in the stage of enhanced convection development.

5. The method according to claim 1, characterized in that, After outputting the strong convective cloud cluster identification result, the method further includes spatiotemporal consistency constraint processing, which includes: Select the first candidate cloud cluster region that was identified as a strong convective cloud cluster at the target observation time. And construct a time window that includes the observation time of the target; Obtain the set of candidate cloud cluster regions and their corresponding strong convection probabilities for each adjacent observation time within the time window, and search for the first candidate cloud cluster region. Matching candidate cloud regions ; Calculate according to the following formula and Spatiotemporal matching cost between : ， In the formula, for and Spatial overlap between them and They are respectively and The position of the center of mass, for The velocity vector of motion, The time difference between the adjacent observation time and the target observation time is the value of the time difference. For normalized distance factor, for and The angle between the directions of motion, For matching weights; When the spatiotemporal matching cost is less than the preset matching threshold, join in The corresponding cross-time association chain; Calculate the duration of the cross-time association chain within the time window. And calculate the cumulative confidence level based on the strong convection probability of each node on the chain. : ， In the formula, For the index of a node in a time-series association chain, The chain length of the chain that is associated across time points. The timestamp corresponding to the time of the target observation. For the first time-related chain The observation timestamps corresponding to each node For the first time-related chain The probability of strong convection at each node This is the time decay coefficient; If the duration Reaching a preset duration threshold and / or cumulative confidence level If the preset threshold is reached, the strong convective cloud cluster identification result of the first candidate cloud cluster region is confirmed as valid, and a continuous early warning signal containing spatiotemporal evolution information is output; otherwise, the strong convective cloud cluster identification result of the first candidate cloud cluster region is marked as instantaneous noise.

6. A system for identifying strong convective cloud clusters based on satellite remote sensing data, characterized in that, The system comprises: steps for implementing the method as described in any one of claims 1-5; The data acquisition unit is used to acquire multi-spectral satellite remote sensing observation data of the area to be monitored within at least one observation period, and to acquire environmental parameter data corresponding to the multi-spectral satellite remote sensing observation data in time and space; and to perform spatiotemporal registration preprocessing on the multi-spectral satellite remote sensing observation data to generate a multi-channel observation data sequence arranged in chronological order. An adaptive candidate extraction unit is used to determine an adaptive dynamic threshold for cloud segmentation based on the environmental parameter data, and to perform candidate extraction processing on the cloud region at the target observation time in the multi-channel observation data sequence using the adaptive dynamic threshold to obtain a set of candidate cloud regions, wherein each candidate cloud region in the set of candidate cloud regions corresponds to a set of spatially connected pixel regions. A physical guidance feature construction unit is used to extract image visual features based on the multi-channel observation data sequence for each candidate cloud region in the candidate cloud region set, and determine environmental physical features based on the environmental parameter data. It then constructs a convection complexity index to characterize the degree of strong convection in the candidate cloud region based on the image visual features and the environmental physical features, and generates a node feature vector for the candidate cloud region. The node feature vector includes the image visual features, the environmental physical features, and the convection complexity index. The graph structure construction unit is used to construct a candidate cloud cluster graph structure based on the candidate cloud cluster region set. The graph nodes of the candidate cloud cluster graph structure correspond to each candidate cloud cluster region, and the graph edges are established between candidate cloud cluster regions that meet the preset distance conditions, which are used to characterize the spatial proximity relationship of the candidate cloud cluster regions. The physical constraint inference unit is used to input the candidate cloud cluster graph structure and the node feature vectors of each graph node into the physical constraint graph convolutional inference model to output the strong convection probability and severity score corresponding to each candidate cloud cluster region, and output the strong convection cloud cluster identification result based on the strong convection probability and the severity score. The strong convection cloud cluster identification result includes at least the geographical range of the strong convection cloud cluster and the corresponding severity score. In the graph convolutional inference process, the physical constraint graph convolutional inference model constrains the aggregation of neighborhood features based on the environmental physical features.

Images