A method for identifying convective nascent by combining multi-source data fusion and deep learning

Abstract

The present application relates to the technical field of meteorological monitoring and short-impending forecast, and particularly relates to a convection initial recognition method combining multi-source data fusion and deep learning. The convection initial recognition method combining multi-source data fusion and deep learning comprises the following steps: S1: obtaining and processing national radar combined reflectivity mosaic data and FY-4B recognition product data to construct a multi-source feature data set for convection initial analysis; S2: obtaining and processing FY-4B full disc data of Shandong region to obtain 1km resolution FY-4B full disc data of Shandong region. The present application constructs a localized feature set by fusing radar and multi-channel satellite data, obtains high-quality samples by combining artificial screening and confidence verification, generates dynamic temperature thresholds suitable for Shandong region, and uses a U-Net model to retrieve night visible light data, so as to finally realize all-weather, high-precision and strong-time-constrained convection initial recognition and short-impending warning.

Description

Technical Field

[0001] This invention relates to the field of meteorological monitoring and short-term forecasting technology, and in particular to a method for identifying convective initiation by combining multi-source data fusion and deep learning. Background Technology

[0002] Convection initiation identification is crucial for short-term severe weather warnings. Current operational identification primarily relies on weather radar reflectivity thresholds, while multi-source data fusion aims to collaboratively utilize heterogeneous radar and satellite observation data. Deep learning, through neural network models, automatically extracts identification features from the data. In recent years, research has demonstrated advantages in improving the timeliness and automation of identification by fusing multi-channel geostationary satellite observations and radar data, and introducing convolutional neural network deep learning methods, thus driving the development of satellite-based identification algorithms.

[0003] However, existing fusion identification methods still have significant limitations. On the one hand, the core algorithms are mostly based on general satellite temperature thresholds, failing to deeply mine and match the localized cloud top characteristics of convective initiation in specific regions, resulting in a high false alarm rate under complex underlying surfaces and weather conditions. On the other hand, such methods heavily rely on visible light channel data, causing the products to fail at night or under low light conditions, making it impossible to achieve continuous all-weather monitoring. This is essentially due to the failure to construct an intelligent discrimination system that closely integrates accurate convective initiation signals from regional radar and possesses spatiotemporal continuity. These limitations reflect the significant deficiencies in the accuracy, precursor detection capability, and all-weather continuity of existing meteorological radar systems in the critical early warning scenario of convective initiation.

[0004] Therefore, there is an urgent need to study an intelligent identification method that can deeply integrate multi-source observation data and effectively improve the radar's accuracy in identifying initial convective signals and the lead time for early warning, so as to overcome the limitations of existing technologies under complex weather and nighttime conditions and improve the ability to capture and predict disastrous convective weather. Summary of the Invention

[0005] To overcome the shortcomings of radar in identifying nascent convection signals, such as insufficient accuracy and time lead, this invention provides a convection nascent identification method that combines multi-source data fusion and deep learning.

[0006] The technical implementation scheme of the present invention is: a convection nascent identification method combining multi-source data fusion and deep learning, comprising the following steps: S1: After acquiring and processing national radar composite reflectivity mosaic data and FY-4B identification product data, a multi-source feature dataset for convection initiation analysis is constructed. S2: Obtain and process the FY-4B full disk data of the Shandong region to obtain the FY-4B full disk data of the Shandong region with a resolution of 1 km. S3: Based on the radar composite reflectivity mosaic data of Shandong region, obtain the radar composite reflectivity mosaic dataset aligned with the latitude and longitude of satellite data of Shandong region, and then filter it to form the weather process dataset required for subsequent analysis. S4: Based on the radar combined reflectivity mosaic data of Shandong region and the weather process dataset, isolated convection points are determined, and the confidence of isolated convection points is filtered to obtain isolated convection initiation points. Based on the isolated convection initiation points, the corresponding FY-4B full disk data of Shandong region is obtained from the multi-source feature dataset and a convection initiation identification training set is generated. S5: The training set for convective primordia recognition is processed to obtain a set of local temperature thresholds that are different from the general thresholds and are specifically for the Shandong region. With the aim of obtaining nighttime visible light channel data, a U-Net deep learning model is constructed and trained to perform data inversion and generate an all-weather convective primordia recognition product.

[0007] Preferably, the step of acquiring and processing nationwide radar composite reflectivity mosaic data and FY-4B identification product data to construct a multi-source feature dataset for convection initiation analysis includes: Spatially crop the national radar composite reflectivity mosaic data to obtain radar composite reflectivity mosaic data for the Shandong region. The latitude and longitude range of the Shandong region is 112° to 120° east longitude and 32° to 41° north latitude. Spatial cropping was performed on the FY-4B identification product data to obtain FY-4B identification product data for the Shandong region; Based on radar composite reflectivity mosaic data and FY-4B identification product data of Shandong region, a multi-source feature dataset for convection initiation analysis is constructed. The multi-source feature dataset contains feature indicators of visible light and infrared multi-channel satellite data. The feature indicators include visible light reflectivity, brightness temperature values ​​and brightness temperature difference values ​​of multiple infrared bands, and the trend of brightness temperature change over time.

[0008] Preferably, the step of acquiring and processing the FY-4B full-disk data of the Shandong region to obtain 1-kilometer resolution FY-4B full-disk data of the Shandong region includes: The resolution of the FY-4B full-disk data for the Shandong region is 4 kilometers. The FY-4B full-disk data of the Shandong region was downscaled using bilinear interpolation to reduce the resolution from 4 kilometers to 1 kilometer, resulting in 1-kilometer resolution FY-4B full-disk data of the Shandong region.

[0009] Preferably, the step of obtaining a radar composite reflectivity mosaic dataset aligned with the latitude and longitude of satellite data for the Shandong region based on radar composite reflectivity mosaic data for the Shandong region, and then filtering it to form a weather process dataset required for subsequent analysis, includes: The data from each time point in the radar composite reflectivity mosaic dataset are displayed graphically. By manually screening, based on preset area thresholds and interference clutter intensity thresholds, weather cases whose strong convective radar echo area exceeds the area threshold and weather cases whose interference clutter intensity exceeds the interference clutter intensity threshold are removed from the weather cases represented by the graphically displayed data. K isolated convective development processes were obtained through manual screening, and these weather cases were used to form the weather process dataset required for subsequent analysis; where K is an integer greater than 1.

[0010] Preferably, the step of determining isolated convection points based on radar composite reflectivity mosaic data of Shandong region combined with weather process dataset, and obtaining isolated convection initiation points by confidence screening of isolated convection points, includes: Based on the chronological order of observation times of radar composite reflectivity mosaic data in Shandong region, data in the weather process dataset are retrieved to locate the distance database with a reflectivity factor greater than 35dBz. For each distance database with a reflectance factor greater than 35 dBz, obtain the reflectance factor data of the current distance database within a 3-kilometer radius of the previous observation time. Determine the presence of data with a reflectivity factor greater than 35 dBz within a 3-kilometer radius of the current storage location: If there are data with a reflectivity factor greater than 35dBz within a 3-kilometer radius of the current distance library, then the point corresponding to the current distance library is determined to be a non-isolated convection point. If there is no data with a reflectivity factor greater than 35dBz within a 3-kilometer radius of the current distance library, then the point corresponding to the current distance library is determined to be an isolated convection initiation point, and the position of the current isolated convection initiation point and the radar echo time are recorded. Based on the isolated convection initiation point and radar echo time, an initiation point confidence screening step is performed.

[0011] Preferably, the step of performing the initial point confidence screening based on the isolated convection initial point and radar echo time includes: For each recorded isolated convection point, continuous time-series observation data of channels C09, C10, C13, and C14 in the 1 km resolution FY-4B full disk data of Shandong region within 30 minutes before the radar echo time are extracted from the multi-source feature dataset used for convection initiation analysis. The extracted continuous time-series observation data is input into a preset initial signal fast filter to calculate the preliminary confidence score of the current isolated convection point; If the initial confidence score exceeds the preset initial screening threshold, the current isolated convection point is marked as a high-confidence isolated convection inception point sample and stored in the high-confidence sample library. If the initial confidence score does not exceed the preset initial screening threshold, the current isolated convection point will be removed.

[0012] Preferably, the step of obtaining the corresponding Shandong region FY-4B full-disk data from the multi-source feature dataset based on isolated convection initiation points and generating a convection initiation recognition training set includes: For each isolated convection initiation point, obtain the 1 km resolution FY-4B full disk data of the Shandong region within 50 minutes before the radar echo time from the multi-source feature dataset used for convection initiation analysis; Based on the FY-4B full-disk data of Shandong region with a resolution of 1 km within 50 minutes prior to the acquired radar echo time, two sets of convection primordial recognition training datasets were extracted and generated: The first training dataset, based on the satellite channels corresponding to the SATCAST algorithm, extracts satellite channel data for each isolated convection initiation point at the radar echo time, 10 minutes in advance, and 20 minutes in advance. The second training dataset extracts satellite channel data with wavelengths outside the visible light and shortwave infrared bands, and extracts satellite channel data for each isolated convection initiation point at the radar echo time, 10 minutes in advance, 20 minutes in advance, and 30 minutes in advance.

[0013] Preferably, the process of processing the convection primordial recognition training set to obtain a set of local temperature thresholds that differ from general thresholds and are specifically applied to the Shandong region includes: Based on the first and second training datasets, the key channel data of Fengyun-4 satellite C02, C04, C09, C10, C13, C14 and C15 included therein were calculated using a preset percentile statistical method to obtain a set of local temperature thresholds that are different from general thresholds and are specifically used for the Shandong region. Construct a convection initiation identification process based on a set of local temperature thresholds that differ from general thresholds and are specifically designed for the Shandong region: Step 1: Obtain the target location data from the C1 to C10 channels of the Fengyun-4 satellite at the radar echo time, 10 minutes in advance, and 20 minutes in advance; Step 2: Based on the set of local temperature thresholds that are different from general thresholds and are specifically used in Shandong region, filter and screen the data from channels C1 to C10. Step 3: Merge the filtered data from each channel into a unified feature set; Step four: Based on the unified feature set, make a judgment. If convection nascent is identified, generate a convection nascent product.

[0014] Preferably, the process of processing the convection primordial recognition training set to obtain a set of local temperature thresholds that differ from general thresholds and are specifically applied to the Shandong region further includes: After generating the initial convection product in step four, the process also includes a cloud-based classification and quality control step for the initial convection product. Acquire Fengyun satellite cloud classification product data that are spatiotemporally matched to the observation time corresponding to the initial convection product for the previous 20 minutes and the previous 10 minutes; the Fengyun satellite cloud classification product data includes type values ​​representing different cloud phases, including at least: water cloud type, supercooled water cloud type, mixed cloud type, ice cloud type, cirrus cloud type and overlapping cloud type; Based on the Fengyun satellite cloud classification product data, areas where the cloud type was either ice cloud or overlapping cloud in the first 20 minutes and the first 10 minutes were selected. The portion of the nascent convection products that matches the selected area is identified as the final nascent convection identification product.

[0015] Preferably, the step of constructing and training a U-Net deep learning model for data inversion with the aim of acquiring nighttime visible light channel data, and generating an all-weather convection initial identification product, includes: During the model training phase, the following steps are performed: Select the 1-kilometer resolution FY-4B full disk data of Shandong region during the daytime, and construct a training dataset containing C09, C10, C12, C13, C14, and C15 channel data as input and C04 channel data as output; The training dataset is subjected to spatiotemporal alignment, resampling, quality control, and normalization. Using the U-Net model structure, regression training was performed with the goal of predicting the reflectance of the C04 channel to obtain the trained U-Net model.

[0016] Preferably, the step of constructing and training a U-Net deep learning model for data inversion with the aim of acquiring nighttime visible light channel data, and generating an all-weather convection initial identification product, includes: During the model application phase, the following steps are performed: acquire 1-kilometer resolution FY-4B full disk data of Shandong region at night, including data from channels C09, C10, C12, C13, C14, and C15; The FY-4B full-disk data of the Shandong region at night with a resolution of 1 km was subjected to spatiotemporal alignment, resampling and normalization processing consistent with the training phase. The processed nighttime 1km resolution Shandong region FY-4B full disk data was input into the trained U-Net model to obtain the inverted predicted C04 channel data. The C04 channel data obtained by inversion is stitched together with the 1 km resolution FY-4B full disk data of Shandong region at night. The complete channel data obtained after splicing is input into a convection initiation identification process based on a set of local temperature thresholds that are different from general thresholds and are specifically used in the Shandong region, to generate an all-weather convection initiation identification product.

[0017] Beneficial Effects: This invention provides an intelligent enhancement scheme for the convective initiation identification capability of meteorological radar systems. By fusing radar combined reflectivity and multi-channel satellite observation data, a multi-source feature dataset of convective initiation focusing on the Shandong region is constructed, and downscaling is used to improve the spatial consistency of the data. Based on high-quality weather process samples selected manually and strict spatiotemporal determination rules for isolated convective initiation points, combined with rapid confidence screening of initiation signals, a training sample set with clear physical meaning is constructed. Furthermore, a dynamic temperature threshold applicable to local cloud top characteristics is generated through statistical learning, breaking through the regional adaptability limitations of general thresholds. Innovatively, the U-Net deep learning model is used to achieve data inversion from infrared and water vapor channels to visible light channels, effectively solving the key bottleneck of missing nighttime identification data. On this basis, the Fengyun satellite cloud classification product is used collaboratively to perform physical verification of the identification results based on cloud top phase state, screening out convective initiation areas with continuous ice phase characteristics, effectively eliminating false signals caused by threshold misjudgment or non-convective systems. Ultimately, by fusing inversion data with multi-channel observations and inputting it into a localized threshold identification process, we achieved all-weather, high-precision, and timely convective initiation identification and short-term early warning capabilities. Attached Figure Description

[0018] Figure 1 This is a flowchart of the convection nascent identification method combining multi-source data fusion and deep learning of the present invention; Figure 2 This is a flowchart of the isolated convection point determination process of the present invention; Figure 3 This is a flowchart illustrating the process of generating the training set for convection nascent recognition in this invention. Figure 4 This is a flowchart of the local temperature threshold acquisition and convection initiation identification process of the present invention; Figure 5 This is a flowchart illustrating the training and application process of the U-Net deep learning model of this invention. Figure 6 This is a flowchart of the initial point confidence screening process for this invention. Detailed Implementation

[0019] The present invention will be further described below with reference to specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0020] Example: A convection nascent identification method combining multi-source data fusion and deep learning, such as... Figures 1-6 As shown, it includes the following steps: S1: After acquiring and processing nationwide radar composite reflectivity mosaic data and FY-4B identification product data, a multi-source feature dataset for convection initiation analysis is constructed, including: Spatially crop the national radar composite reflectivity mosaic data to obtain radar composite reflectivity mosaic data for the Shandong region. The latitude and longitude range of the Shandong region is 112° to 120° east longitude and 32° to 41° north latitude. Spatial cropping was performed on the FY-4B identification product data to obtain FY-4B identification product data for the Shandong region; Based on radar composite reflectivity mosaic data and FY-4B identification product data of Shandong region, a multi-source feature dataset for convection initiation analysis is constructed. The multi-source feature dataset contains feature indicators of visible light and infrared multi-channel satellite data. The feature indicators include visible light reflectivity, brightness temperature values ​​and brightness temperature difference values ​​of multiple infrared bands, and the trend of brightness temperature change over time.

[0021] It should be noted that the initial steps of this invention aim to simultaneously acquire official identification products from the FY-4B satellite, covering the entire country, using radar composite reflectivity mosaics. The radar mosaics provide direct detection signals of convection intensity; the full satellite data contains multispectral information that can invert cloud top physical states; and both identification products integrate feature indicators generated by previous algorithm processing. To address the shortcomings of general algorithms in adapting to specific geographical and climatic conditions, this invention uniformly cropped the aforementioned nationwide data to the Shandong region, thereby accurately defining the analysis scope and improving computational efficiency. The resulting radar and satellite dataset for the Shandong region lays the foundation for subsequently building models tailored to local characteristics.

[0022] After integrating the two types of regionalized data mentioned above, a multi-source feature dataset specifically for convective initiation analysis was constructed through rigorous spatiotemporal matching and fusion processing. The core of this dataset is the extraction and integration of multi-channel satellite feature indicators: visible light reflectance is used to identify cloud contours; brightness temperature values ​​in multiple infrared bands directly indicate cloud top height and temperature; brightness temperature differences between different bands help distinguish cloud phases and development stages; and the trend of brightness temperature over time dynamically reflects the vertical development intensity of cloud clusters—a larger trend usually indicates a more intense cloud top uplift or cooling process, while a smaller trend indicates relatively stable cloud layers. These fused multi-dimensional features collectively constitute a key information system for characterizing the initiation process of convective clouds, providing reliable data support for establishing a high-precision, localized identification model. To eliminate scale differences in visible light reflectance and infrared brightness temperature features with different dimensions during fusion and subsequent analysis, the feature indicators need to be standardized after constructing the multi-source feature dataset. For example, max-min normalization or Z-score standardization methods can be used to convert the feature values ​​to a uniform dimension or a similar numerical range.

[0023] It should be noted that the letter "C" before the satellite channel name in this article is the abbreviation for "Channel". For example, "C04" represents channel 04 (visible light channel) of Fengyun-4 satellite, and "C13" represents channel 13 (infrared channel). This abbreviation will be used to refer to the channel in the following text.

[0024] S2: After acquiring and processing the FY-4B full-disk data for the Shandong region, we obtain 1-kilometer resolution FY-4B full-disk data for the Shandong region, including: The resolution of the FY-4B full-disk data for the Shandong region is 4 kilometers. The FY-4B full-disk data of the Shandong region was downscaled using bilinear interpolation to reduce the resolution from 4 kilometers to 1 kilometer, resulting in 1-kilometer resolution FY-4B full-disk data of the Shandong region.

[0025] It should be noted that the FY-4B full-disk data for the Shandong region was obtained through a national-level satellite data receiving system. This data includes multispectral channel observations used to retrieve cloud top characteristics such as brightness temperature and reflectivity. The original data has a resolution of 4 kilometers, which limits spatial detail. Therefore, bilinear interpolation was used for downscaling to refine the resolution to 1 kilometer. Bilinear interpolation calculates the weighted average of neighboring pixels, which smoothly generates intermediate values, thus maintaining the spatial continuity of the data. After the resolution was improved to 1 kilometer, the satellite observations could be scaled together with the high spatial accuracy radar reflectivity data. This enhances the consistency of multi-source data fusion, improves the detection capability of small, rapidly evolving weather systems in the identification of convective initiation, and effectively addresses the problem of insufficient identification accuracy caused by data resolution mismatch in existing technologies.

[0026] S3: Based on the radar composite reflectivity mosaic data for the Shandong region, obtain a radar composite reflectivity mosaic dataset aligned with the latitude and longitude of the satellite data for the Shandong region, and then filter it to form the weather process dataset required for subsequent analysis, including: The data from each time point in the radar composite reflectivity mosaic dataset are displayed graphically. By manually screening, based on preset area thresholds and interference clutter intensity thresholds, weather cases whose strong convective radar echo area exceeds the area threshold and weather cases whose interference clutter intensity exceeds the interference clutter intensity threshold are removed from the weather cases represented by the graphically displayed data. K isolated convective development processes were obtained through manual screening, and these weather cases were used to form the weather process dataset required for subsequent analysis; where K is an integer greater than 1.

[0027] It's important to note that the core of this step lies in building a high-purity, targeted sample library for subsequent model training. First, the radar composite reflectivity mosaic data for the Shandong region must be strictly aligned with the satellite data using latitude and longitude grids. This is because radar and satellite are heterogeneous observation systems; only by unifying the spatial reference can subsequent point-to-point data fusion and feature correlation have physical meaning. Visualizing all time-series data essentially converts the digital sequence into visually interpretable weather process animations, allowing experts to intuitively identify them using meteorological knowledge. For example, a continuous animation clearly shows the isolated development process of an echo patch from nothing to its rapid intensification.

[0028] Setting screening thresholds requires consideration of radar meteorology principles and local climate statistics. The area threshold aims to eliminate organized mesoscale convective systems, using the typical maximum horizontal scale of isolated nascent convective cells as a reference upper limit based on historical case statistics. The interference clutter intensity threshold filters out non-meteorological echoes, using the statistically high value of the long-term echo intensity distribution of fixed objects around the radar station as the judgment boundary. Therefore, the "preset area threshold and interference clutter intensity threshold" are values ​​determined after quantitative analysis of historical data based on the aforementioned statistical principles. During manual screening, experts use these quantitative thresholds for judgment, ensuring the consistency and repeatability of sample screening standards. The purpose of manual screening is to proactively eliminate two types of samples unsuitable for "nascent" studies: large-scale mature convection provides cloud top characteristics drastically different from the nascent stage; strong interference clutter is completely irrelevant noise. Retaining them would severely pollute the training set, preventing the model from focusing on learning the core physical signals of nascent convection, thus exacerbating the false alarm problem prevalent in existing algorithms.

[0029] The K isolated convection cases selected at the end constitute a "clean" dataset representing typical local primordial processes. This dataset serves as a high-quality training set for subsequent machine learning and forms the basis for developing high-precision, low-false-positive localized recognition algorithms, directly addressing the problem of poor adaptability of general-purpose algorithms due to inaccurate training samples.

[0030] S4-1: Based on radar composite reflectivity mosaic data of Shandong region combined with weather process dataset, isolated convection points are identified, and confidence screening of isolated convection points is performed to obtain isolated convection initiation points, including: Based on the chronological order of observation times of radar composite reflectivity mosaic data in Shandong region, data in the weather process dataset are retrieved to locate the distance database with a reflectivity factor greater than 35dBz. For each distance database with a reflectance factor greater than 35 dBz, obtain the reflectance factor data of the current distance database within a 3-kilometer radius of the previous observation time. Determine the presence of data with a reflectivity factor greater than 35 dBz within a 3-kilometer radius of the current storage location: If there are data with a reflectivity factor greater than 35dBz within a 3-kilometer radius of the current distance library, then the point corresponding to the current distance library is determined to be a non-isolated convection point. If there is no data with a reflectivity factor greater than 35dBz within a 3-kilometer radius of the current distance library, then the point corresponding to the current distance library is determined to be an isolated convection initiation point, and the position of the current isolated convection initiation point and the radar echo time are recorded. It should be noted that the reference Figure 2 To accurately pinpoint the exact moment and location of convective initiation from the filtered weather process data, this step performs a point-by-point analysis based on the spatiotemporal evolution of radar reflectivity. The weather process dataset is retrieved in chronological order of observation times because convective initiation is a process with a clear temporal orientation; therefore, a retrospective analysis following its natural development timeline is necessary to distinguish the "initiation" moment from the preceding "development" state. The retrieval process involves reading the data files sequentially and scanning the radar mosaic for each time point.

[0031] The core of the analysis is identifying the location where the reflectivity factor first reaches a specific intensity threshold. The reflectivity factor, measured in dBz, quantitatively characterizes the ability of atmospheric precipitation particles to scatter radar signals. For example, a reflectivity factor of 35 dBz typically indicates the presence of strong precipitation particles and is a commonly used criterion for convective activity reaching a certain intensity. The range library is the smallest spatial analysis unit in radar polar coordinate data; each library represents a sampling volume at a specific direction and range.

[0032] For each range with a reflectivity factor exceeding 35 dBz, the algorithm traces the echo situation within a 3-kilometer radius of the previous observation time. This 3-kilometer radius balances the spatial scale of a typical newly formed convective cell with the spatial resolution of radar data. If a strong echo exceeding 35 dBz is detected within this range in the previous observation time, it indicates that the current strong echo is caused by the movement or expansion of an existing convective system, thus determining that the point is not a newly formed "isolated convection initiation point." Conversely, if no strong echo is detected within a 3-kilometer radius in the previous observation time, it means that the current strong echo developed rapidly locally within a short period, meeting the physical definition of "isolated convection initiation."

[0033] Once identified as an isolated convection initiation point, its geographical location (latitude and longitude coordinates) and radar echo time must be accurately recorded. The geographical location is used for spatial matching with satellite data to extract multi-channel cloud top features corresponding to that point; the radar echo time serves as the "baseline time" for convection initiation, used to extract cloud top state evolution sequences with different lead times prior to that moment from historical satellite data. This step, by establishing a precise spatiotemporal correlation between the "time of strong signal appearance" detected by radar and the "early feature evolution" observed by satellite, provides a crucial anchor point for subsequent training of intelligent models capable of identifying early signs of initiation, directly addressing the problem of insufficient spatiotemporal correlation mining of initiation signals in existing algorithms.

[0034] Based on isolated convection points and radar echo times, a primary point confidence screening step is performed, including: For each recorded isolated convection point, continuous time-series observation data of channels C09, C10, C13, and C14 in the 1 km resolution FY-4B full disk data of Shandong region within 30 minutes before the radar echo time are extracted from the multi-source feature dataset used for convection initiation analysis. The extracted continuous time-series observation data is input into a preset initial signal fast filter to calculate the preliminary confidence score of the current isolated convection point; If the initial confidence score exceeds the preset initial screening threshold, the current isolated convection point is marked as a high-confidence isolated convection inception point sample and stored in the high-confidence sample library. If the initial confidence score does not exceed the preset initial screening threshold, the current isolated convection initiation point will be removed.

[0035] It should be noted that the reference Figure 6Although the spatiotemporal evolution of radar reflectivity can locate potential isolated convection points, some of these points are unreliable due to radar clutter, terrain obstruction, or weak precipitation cells. Therefore, confidence screening is necessary. The core principle of this screening is to use the typical evolution patterns of cloud top physical characteristics observed by satellite during the early stages of cloud formation to verify the physical plausibility of the "initial signals" detected by radar.

[0036] For each isolated convection point, continuous time-series data from FY-4B satellite channels C09, C10, C13, and C14 were extracted within 30 minutes prior to the radar echo time. This 30-minute period represents the critical developmental stage of convective clouds from their initial formation to radar detectability. For example, the continuous decrease in brightness temperature in channel C13 (10.8 μm) directly reflects the uplift and cooling process of the cloud top. Continuous time-series observation data refers to the numerical sequence of these four channels at that point location over multiple consecutive observation times (e.g., every 5 minutes).

[0037] The pre-defined rapid filter for nascent signals is a discriminant function constructed based on prior knowledge. This "prior knowledge" originates from meteorological satellite remote sensing and convective cloud physics, and its core is to identify key satellite observation feature patterns related to convective initiation, such as the continuous decrease in infrared brightness temperature caused by developing convective cloud tops and the characteristic evolution of brightness-temperature differences between specific channels. The "pre-defined" nature of this filter means that its internal discriminant logic and calculation method are determined before being applied to this method. Inputting the extracted continuous time-series observation data into this filter calculates a quantitative value characterizing the degree of agreement between the satellite's early-stage features and the convective initiation pattern at that point—the preliminary confidence score.

[0038] The preset initial screening threshold is a benchmark value used to evaluate the level of the aforementioned preliminary confidence score. The specific value of this threshold is determined through statistical analysis of an independent database of historically confirmed high-quality nascent cases. Specifically, the filter is first used to calculate the preliminary confidence score of all samples in this historical database; then, the distribution of these scores is analyzed, and a suitable statistic (e.g., the median or 75th percentile of the score distribution) is selected as the preset initial screening threshold. This method ensures that the threshold setting objectively reflects the general characteristic level of historical high-quality nascent events.

[0039] If the initial confidence score of an isolated convection point exceeds the preset screening threshold, it is marked as a high-confidence isolated convection initiation point sample and stored in the high-confidence sample database. If the score does not exceed the threshold, it is removed. The principle is that points with scores below the threshold do not conform to typical convection initiation patterns in their early satellite characteristic evolution, corresponding to misjudgments or weak development processes. Removing them avoids introducing noise into the subsequent training set. This screening step essentially establishes a strong correlation between the "signal appearance" detected by radar and the "physical process evolution" observed by satellite, effectively improving the physical consistency and reliability of the training samples, and directly addressing the core problem of insufficient initiation signal discrimination accuracy mentioned in the background technology.

[0040] S4-2: Based on isolated convection initiation points, obtain the corresponding Shandong region FY-4B full-disk data from the multi-source feature dataset and generate a convection initiation recognition training set, including: For each isolated convection initiation point, obtain the 1 km resolution FY-4B full disk data of the Shandong region within 50 minutes before the radar echo time from the multi-source feature dataset used for convection initiation analysis; Based on the FY-4B full-disk data of Shandong region with a resolution of 1 km within 50 minutes prior to the acquired radar echo time, two sets of convection primordial recognition training datasets were extracted and generated: The first training dataset, based on the satellite channels corresponding to the SATCAST algorithm, extracts satellite channel data for each isolated convection initiation point at the radar echo time, 10 minutes in advance, and 20 minutes in advance. The second training dataset extracts satellite channel data with wavelengths outside the visible light and shortwave infrared bands, and extracts satellite channel data for each isolated convection initiation point at the radar echo time, 10 minutes in advance, 20 minutes in advance, and 30 minutes in advance.

[0041] It should be noted that the reference Figure 3 The core task of this step is to utilize the confirmed high-quality convection initiation points to construct a satellite feature sequence dataset that can characterize their development process, providing input for subsequent threshold calculations and model training. Extracting 1-kilometer resolution satellite data within 50 minutes before the radar echo time is based on the time scale consideration of the physical development process of convective clouds; the 50-minute window is sufficient to cover the complete early stage from the initial formation of the cloud system to the first detection of strong echoes by the radar, while the 1-kilometer resolution matches the radar data scale, ensuring the spatial precision of feature extraction.

[0042] Two training datasets were generated to serve different algorithm development purposes. The first dataset is based on the mature SATCAST algorithm framework (SATCAST is a convection initiation identification algorithm based on multi-channel brightness temperature changes of geostationary satellites). This algorithm establishes a set of empirical rules for identifying convection initiation using multi-channel brightness temperature and its changes. Data from radar echo times, 10 minutes in advance, and 20 minutes in advance were extracted to construct a feature snapshot containing the "initiation moment" and two key advance times, enabling the new algorithm to learn the typical feature patterns of the channels of interest to SATCAST in the initiation imminent stage. The second dataset expands the data dimensions by specifically extracting mid- and long-wave infrared channel data beyond the visible and short-wave infrared bands. Specifically, in the FY-4B satellite data utilized in this invention, the "satellite channels with wavelengths outside the visible light and shortwave infrared bands" refer to the mid- and long-wave infrared channels they contain, such as C09 (10.4 μm), C10 (7.3 μm), C12 (10.8 μm), C13 (10.8 μm), C14 (11.2 μm), and C15 (12.3 μm). These channels are less affected by sunlight and contain information on cloud height and phase. Data from four time points, including 30 minutes in advance, is extracted to capture changes in cloud top thermal characteristics that develop earlier and evolve more slowly, providing training material for developing new algorithms that are independent of visible light and have longer-term warning potential. By constructing these two datasets—one inheriting the classic framework and the other exploring new features—this method lays a solid data foundation for developing reliable and forward-looking localized identification algorithms.

[0043] S5-1: Processing the convection nascent recognition training set yields a set of local temperature thresholds that differ from general thresholds and are specifically designed for the Shandong region, including: Based on the first and second training datasets, the key channel data of Fengyun-4 satellite C02, C04, C09, C10, C13, C14 and C15 included therein were calculated using a preset percentile statistical method to obtain a set of local temperature thresholds that are different from general thresholds and are specifically used for the Shandong region. Construct a convection initiation identification process based on a set of local temperature thresholds that differ from general thresholds and are specifically designed for the Shandong region: Step 1: Obtain the target location data from the C1 to C10 channels of the Fengyun-4 satellite at the radar echo time, 10 minutes in advance, and 20 minutes in advance; Step 2: Based on the set of local temperature thresholds that are different from general thresholds and are specifically used in Shandong region, filter and screen the data from channels C1 to C10. Step 3: Merge the filtered data from each channel into a unified feature set; Step four: Based on the unified feature set, make a judgment. If convection nascent is identified, generate a convection nascent product.

[0044] It should be noted that the reference Figure 4 This step aims to statistically generate a set of temperature thresholds for identifying the initial stage of convection in the Shandong region based on localized training samples, and to construct a corresponding automated identification process. The reason for using preset percentiles (such as the 80th percentile) for the CO2 and CO4 key channel data is that this method can effectively characterize the typical distribution characteristics of satellite observations in local convection initiation cases, thus obtaining robust and regionally representative thresholds. This directly overcomes the problem of insufficient adaptability caused by the general threshold ignoring geographical and climatic differences.

[0045] The standard procedure for calculating the local temperature threshold using a statistical method based on preset percentiles is as follows: 1. Data Collection: Extract the brightness temperature observation values ​​of a specific channel (e.g., C13) from all high-confidence samples in the training dataset at a specific time (e.g., radar echo time), forming a sample data list for that channel. 2. Sorting: Sort the sample data list in ascending order of value. 3. Percentile Positioning: Calculate the position index of the channel in the sorted list based on a preset percentile P (e.g., 80). The calculation method is common knowledge in the field, for example: position = P% * (total number of samples + 1). 4. Threshold Determination: Based on the calculated position index (integer or decimal), determine the brightness temperature value corresponding to that position using a conventional linear interpolation method. This value is the local temperature threshold for that channel. 5. Iterative Calculation: Repeat steps 1 to 4 for each channel requiring a threshold (C02, C04, C09, C10, C13, C14, C15) to obtain a complete set of local temperature thresholds.

[0046] The constructed identification process first acquires C1 to C10 channel data for the target location at three time points: the radar echo time, 10 minutes prior, and 20 minutes prior. These three time points cover the critical evolution period of the initial stage of convection, and the multi-channel data together constitute the characteristic basis for diagnosing the dynamic changes in cloud top status. Subsequently, the data of each channel is filtered and screened using a local temperature threshold. This step is essentially an initial feature selection based on local experience of the original observations, eliminating noise or irrelevant data that clearly does not conform to the initial characteristics. Without this screening, a large amount of irrelevant information would interfere with subsequent judgments. Next, the filtered multi-channel, multi-time data is fused into a unified feature vector. This feature set comprehensively represents the multi-dimensional attributes of the cloud top of the target point in the initial stage. Specifically, the observation values ​​(such as brightness temperature and reflectivity) of each time point and each filtered channel are concatenated in a preset order (e.g., by time sequence, then by channel number) to form a one-dimensional numerical array, which constitutes the unified feature vector. Finally, based on this feature set, judgment rules (such as multi-indicator joint identification) are applied. If the conditions are met, the location is marked as a convection initiation and the corresponding product is generated. This entire process realizes the automated and localized conversion from multi-time satellite observations to initiation products, significantly improving the accuracy and reliability of identification.

[0047] The training set for convection primordial recognition was processed to obtain a set of local temperature thresholds that differ from general thresholds and are specifically designed for the Shandong region. This also includes: After generating the initial convection product in step four, the process also includes a cloud-based classification and quality control step for the initial convection product. Acquire Fengyun satellite cloud classification product data that are spatiotemporally matched to the observation time corresponding to the initial convection product for the previous 20 minutes and the previous 10 minutes; the Fengyun satellite cloud classification product data includes type values ​​representing different cloud phases, including at least: water cloud type, supercooled water cloud type, mixed cloud type, ice cloud type, cirrus cloud type and overlapping cloud type; Based on the Fengyun satellite cloud classification product data, areas where the cloud type was either ice cloud or overlapping cloud in the first 20 minutes and the first 10 minutes were selected. The portion of the nascent convection products that matches the selected area is identified as the final nascent convection identification product.

[0048] It should be noted that, to enhance the reliability of convective cloud formation products based on the collaborative identification of radar and satellite data, this invention introduces a quality control step based on Fengyun satellite cloud classification products (CLT) after initial judgment using localized temperature thresholds. This step aims to utilize the physical information of cloud top phases to filter the physical plausibility of the initial identification results. Specifically, the cloud classification product is an L2-level quantitative product released by the National Satellite Meteorological Center. By fusing multi-channel satellite observations, each pixel is divided into six main cloud phase types, including water clouds, supercooled water clouds, mixed clouds, ice clouds, cirrus clouds, and overlapping clouds. Among them, ice clouds and overlapping clouds indicate the presence of opaque and relatively dense ice crystal particles at the cloud top. These clouds are usually well-developed, with high cloud tops, consistent with the typical cloud top characteristics of the initial stage of convective clouds. This method extracts cloud classification data that precisely matches the initial judgment time in the 20 minutes and 10 minutes prior to the initial judgment time, retaining only areas where the cloud type is determined to be ice clouds or overlapping clouds in both consecutive time periods. This requirement is based on the physical laws of convective cloud development: convective cells that can rapidly develop and be detected by radar within approximately half an hour should have cloud tops exhibiting persistent icy phase characteristics in previous observations. Therefore, spatially matching and filtering the initially identified convective initiation products with the areas verified by persistent icy phase cloud tops essentially involves using independent cloud physics observations to verify the physical consistency of radar detection signals. This effectively filters out false initiation signals caused by threshold misjudgments or other non-convective weather systems, thereby significantly improving the accuracy and reliability of the final convective initiation identification products. This directly addresses the problem of high false alarm rates in existing radar identification technologies due to insufficient single criteria.

[0049] S5-2: To acquire nighttime visible light channel data, construct and train a U-Net deep learning model for data inversion and generate an all-weather convection initial recognition product, including: During the model training phase, the following steps are performed: Select the 1-kilometer resolution FY-4B full disk data of Shandong region during the daytime, and construct a training dataset containing C09, C10, C12, C13, C14, and C15 channel data as input and C04 channel data as output; The training dataset is subjected to spatiotemporal alignment, resampling, quality control, and normalization. Using the U-Net model structure, regression training was performed with the goal of predicting the reflectance of the C04 channel to obtain the trained U-Net model.

[0050] It should be noted that the reference Figure 5To overcome the bottleneck of existing algorithms failing at night due to the lack of visible light data, this step aims to train a deep neural network to achieve the inversion from multi-channel infrared data to visible light channels. The model training uses 1km resolution FY-4B data from daytime because it simultaneously provides usable C04 visible light channel data and C09 and C10 infrared channel data, enabling the construction of a mapping relationship from "infrared feature combination" to "visible light reflectance" as supervised learning samples. The selection of specific infrared and water vapor channels C09, C10, C12, C13, C14, and C15 as inputs is based on their physical characteristics; they carry information about cloud top temperature, altitude, and phase, which are physically related to the optical reflectance characteristics of clouds. The training dataset undergoes spatiotemporal alignment, resampling, quality control, and normalization to eliminate spatiotemporal misalignments and dimensional differences in observations, ensuring the physical consistency and numerical stability of the input data, and providing a clean and standardized data foundation for model learning. Regression training utilizes encoder-decoder structures like U-Net (a convolutional neural network with an encoder-decoder architecture suitable for image regression tasks) to fully leverage its ability to capture multi-scale spatial contextual information, thereby accurately predicting the C04 channel reflectance of each pixel. The core function of the final trained model is to intelligently "fill in" missing visible light data at night using continuously available infrared channel observations, thus providing crucial input for subsequent all-weather recognition processes based on multi-channel thresholds.

[0051] During the model application phase, the following steps are performed: acquire 1-kilometer resolution FY-4B full disk data of Shandong region at night, including data from channels C09, C10, C12, C13, C14, and C15; The FY-4B full-disk data of the Shandong region at night with a resolution of 1 km was subjected to spatiotemporal alignment, resampling and normalization processing consistent with the training phase. The processed nighttime 1km resolution Shandong region FY-4B full disk data was input into the trained U-Net model to obtain the inverted predicted C04 channel data. The C04 channel data obtained by inversion is stitched together with the 1 km resolution FY-4B full disk data of Shandong region at night. The complete channel data obtained after splicing is input into a convection initiation identification process based on a set of local temperature thresholds that are different from general thresholds and are specifically used in the Shandong region, to generate an all-weather convection initiation identification product.

[0052] It should be noted that the reference Figure 5To achieve continuous, all-weather monitoring of initial convection formation, this step focuses on addressing the core bottleneck of missing nighttime visible light data. When applying the model, firstly, 1km resolution FY-4B full-disk data is acquired at night. The C09, C10, C12, C13, C14, and C15 infrared and water vapor channels remain valid even in the absence of light, continuously providing information on cloud top thermal status. This data must undergo spatiotemporal alignment, resampling, and normalization preprocessing identical to that used in the training phase to eliminate differences in the observation system and ensure that its numerical distribution strictly matches the basic features learned during model training. This is a prerequisite for the model to make reliable extrapolation predictions.

[0053] The preprocessed nighttime multi-channel data is then input into the trained U-Net model. This model utilizes deep nonlinear mapping relationships learned from daytime data to intelligently estimate the reflectance value of the missing C04 channel from infrared features. After obtaining the retrieved C04 data, it needs to be fused with the original nighttime infrared data into a whole through channel-dimensional concatenation. Specifically, the retrieved C04 data matrix (two-dimensional spatial field) is treated as a new independent channel layer and combined with the original C09 to C15 channel data (together forming a three-dimensional data cube) along the channel dimension in terms of data structure. For example, array manipulation functions are used to append the C04 data layer to the original data cube, thereby generating a new dataset with a complete structure containing 7 channels.

[0054] The resulting complete channel dataset, in terms of data dimensions and information composition, is equivalent to the available daytime data. This dataset is then fed into a discrimination process based on localized temperature thresholds. Since all necessary channel information is now available, the process can perform comprehensive identification normally at night, ultimately outputting a convection priming identification product consistent with the daytime mechanism (e.g., a spatial distribution map indicating the probability of nighttime priming). This process fundamentally overcomes the absolute dependence of traditional methods on visible light data, achieving uninterrupted, all-weather monitoring capabilities.

[0055] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A convection nascent identification method combining multi-source data fusion and deep learning, characterized in that, Includes the following steps: S1: After acquiring and processing national radar composite reflectivity mosaic data and FY-4B identification product data, a multi-source feature dataset for convection initiation analysis is constructed. S2: Obtain and process the FY-4B full disk data of the Shandong region to obtain the FY-4B full disk data of the Shandong region with a resolution of 1 km. S3: Based on the radar composite reflectivity mosaic data of Shandong region, obtain the radar composite reflectivity mosaic dataset aligned with the latitude and longitude of satellite data of Shandong region, and then filter it to form the weather process dataset required for subsequent analysis. S4: Based on the radar combined reflectivity mosaic data of Shandong region and the weather process dataset, isolated convection points are determined, and the confidence of isolated convection points is filtered to obtain isolated convection initiation points. Based on the isolated convection initiation points, the corresponding FY-4B full disk data of Shandong region is obtained from the multi-source feature dataset and a convection initiation identification training set is generated. S5: The training set for convective primordia recognition is processed to obtain a set of local temperature thresholds that are different from the general thresholds and are specifically for the Shandong region. With the aim of obtaining nighttime visible light channel data, a U-Net deep learning model is constructed and trained to perform data inversion and generate an all-weather convective primordia recognition product.

2. The convection nascent identification method combining multi-source data fusion and deep learning as described in claim 1, characterized in that, The process of acquiring and processing nationwide radar composite reflectivity mosaic data and FY-4B identification product data to construct a multi-source feature dataset for convection initiation analysis includes: Spatially crop the national radar composite reflectivity mosaic data to obtain radar composite reflectivity mosaic data for the Shandong region. The latitude and longitude range of the Shandong region is 112° to 120° east longitude and 32° to 41° north latitude. Spatial cropping was performed on the FY-4B identification product data to obtain FY-4B identification product data for the Shandong region; Based on radar composite reflectivity mosaic data and FY-4B identification product data of Shandong region, a multi-source feature dataset for convection initiation analysis is constructed. The multi-source feature dataset contains feature indicators of visible light and infrared multi-channel satellite data. The feature indicators include visible light reflectivity, brightness temperature values ​​and brightness temperature difference values ​​of multiple infrared bands, and the trend of brightness temperature change over time.

3. The convection nascent identification method combining multi-source data fusion and deep learning as described in claim 1, characterized in that, The process of acquiring and processing the full-disk FY-4B data for the Shandong region to obtain 1-kilometer resolution full-disk FY-4B data for the Shandong region includes: The resolution of the FY-4B full-disk data for the Shandong region is 4 kilometers. The FY-4B full-disk data of the Shandong region was downscaled using bilinear interpolation to reduce the resolution from 4 kilometers to 1 kilometer, resulting in 1-kilometer resolution FY-4B full-disk data of the Shandong region.

4. The convection nascent identification method combining multi-source data fusion and deep learning as described in claim 1, characterized in that, The process involves obtaining a radar composite reflectivity mosaic dataset aligned with the latitude and longitude of satellite data for the Shandong region based on radar composite reflectivity mosaic data for the Shandong region, and then filtering it to form a weather process dataset required for subsequent analysis, including: The data from each time point in the radar composite reflectivity mosaic dataset are displayed graphically. By manually screening, based on preset area thresholds and interference clutter intensity thresholds, weather cases whose strong convective radar echo area exceeds the area threshold and weather cases whose interference clutter intensity exceeds the interference clutter intensity threshold are removed from the weather cases represented by the graphically displayed data. K isolated convective development processes were obtained through manual screening, and these weather cases were used to form the weather process dataset required for subsequent analysis; where K is an integer greater than 1.

5. The convection nascent identification method combining multi-source data fusion and deep learning according to claim 1, characterized in that, The process of identifying isolated convection points based on radar composite reflectivity mosaic data of the Shandong region combined with weather process datasets, and then performing confidence screening on these isolated convection points to obtain the initial points of isolated convection, includes: Based on the chronological order of observation times of radar composite reflectivity mosaic data in Shandong region, data in the weather process dataset are retrieved to locate the distance database with a reflectivity factor greater than 35dBz. For each distance database with a reflectance factor greater than 35 dBz, obtain the reflectance factor data of the current distance database within a 3-kilometer radius of the previous observation time. Determine the presence of data with a reflectivity factor greater than 35 dBz within a 3-kilometer radius of the current storage location: If there are data with a reflectivity factor greater than 35dBz within a 3-kilometer radius of the current distance library, then the point corresponding to the current distance library is determined to be a non-isolated convection point. If there is no data with a reflectivity factor greater than 35dBz within a 3-kilometer radius of the current distance library, then the point corresponding to the current distance library is determined to be an isolated convection initiation point, and the position of the current isolated convection initiation point and the radar echo time are recorded. Based on the isolated convection initiation point and radar echo time, an initiation point confidence screening step is performed.

6. The convection nascent identification method combining multi-source data fusion and deep learning according to claim 5, characterized in that, The step of performing initial point confidence screening based on isolated convection initiation points and radar echo times includes: For each recorded isolated convection point, continuous time-series observation data of channels C09, C10, C13, and C14 in the 1 km resolution FY-4B full disk data of Shandong region within 30 minutes before the radar echo time are extracted from the multi-source feature dataset used for convection initiation analysis. The extracted continuous time-series observation data is input into a preset initial signal fast filter to calculate the preliminary confidence score of the current isolated convection point; If the initial confidence score exceeds the preset initial screening threshold, the current isolated convection point is marked as a high-confidence isolated convection inception point sample and stored in the high-confidence sample library. If the initial confidence score does not exceed the preset initial screening threshold, the current isolated convection point will be removed.

7. A convection nascent identification method combining multi-source data fusion and deep learning according to claim 1, characterized in that, The process of obtaining the corresponding FY-4B full-disk data of the Shandong region from a multi-source feature dataset based on isolated convection initiation points and generating a convection initiation recognition training set includes: For each isolated convection initiation point, obtain the 1 km resolution FY-4B full disk data of the Shandong region within 50 minutes before the radar echo time from the multi-source feature dataset used for convection initiation analysis; Based on the FY-4B full-disk data of Shandong region with a resolution of 1 km within 50 minutes prior to the acquired radar echo time, two sets of convection primordial recognition training datasets were extracted and generated: The first training dataset, based on the satellite channels corresponding to the SATCAST algorithm, extracts satellite channel data for each isolated convection initiation point at the radar echo time, 10 minutes in advance, and 20 minutes in advance. The second training dataset extracts satellite channel data with wavelengths outside the visible light and shortwave infrared bands, and extracts satellite channel data for each isolated convection initiation point at the radar echo time, 10 minutes in advance, 20 minutes in advance, and 30 minutes in advance.

8. The convection nascent identification method combining multi-source data fusion and deep learning according to claim 1, characterized in that, The convection primordial recognition training set is processed to obtain a set of local temperature thresholds that differ from general thresholds and are specifically used in the Shandong region, including: Based on the first and second training datasets, the key channel data of Fengyun-4 satellite 02, C04, C09, C10, C13, C14 and C15 included therein were calculated using a preset percentile statistical method to obtain a set of local temperature thresholds that are different from general thresholds and are specifically used for the Shandong region. Construct a convection initiation identification process based on a set of local temperature thresholds that differ from general thresholds and are specifically designed for the Shandong region: Step 1: Obtain the target location data from the C1 to C10 channels of the Fengyun-4 satellite at the radar echo time, 10 minutes in advance, and 20 minutes in advance; Step 2: Based on the set of local temperature thresholds that are different from general thresholds and are specifically used in Shandong region, filter and screen the data from channels C1 to C10. Step 3: Merge the filtered data from each channel into a unified feature set; Step four: Based on the unified feature set, make a judgment. If convection nascent is identified, generate a convection nascent product.

9. A convection nascent identification method combining multi-source data fusion and deep learning according to claim 8, characterized in that, The process of processing the initial convection recognition training set to obtain a set of local temperature thresholds that differ from general thresholds and are specifically designed for the Shandong region also includes: After generating the initial convection product in step four, the process also includes a cloud-based classification and quality control step for the initial convection product. Acquire Fengyun satellite cloud classification product data that are spatiotemporally matched to the observation time corresponding to the initial convection product for the previous 20 minutes and the previous 10 minutes; the Fengyun satellite cloud classification product data includes type values ​​representing different cloud phases, including at least: water cloud type, supercooled water cloud type, mixed cloud type, ice cloud type, cirrus cloud type and overlapping cloud type; Based on the Fengyun satellite cloud classification product data, areas where the cloud type was either ice cloud or overlapping cloud in the first 20 minutes and the first 10 minutes were selected. The portion of the nascent convection products that matches the selected area is identified as the final nascent convection identification product.

10. A convection nascent identification method combining multi-source data fusion and deep learning according to claim 1, characterized in that, The purpose of acquiring nighttime visible light channel data is to construct and train a U-Net deep learning model for data inversion and generate an all-weather convection initial recognition product, including: During the model training phase, the following steps are performed: Select the 1-kilometer resolution FY-4B full disk data of Shandong region during the daytime, and construct a training dataset containing C09, C10, C12, C13, C14, and C15 channel data as input and C04 channel data as output; The training dataset is subjected to spatiotemporal alignment, resampling, quality control, and normalization. Using the U-Net model structure, regression training was performed with the goal of predicting the reflectance of the C04 channel to obtain the trained U-Net model.

11. A convection nascent identification method combining multi-source data fusion and deep learning according to claim 1, characterized in that, The purpose of acquiring nighttime visible light channel data is to construct and train a U-Net deep learning model for data inversion and generate an all-weather convection initial recognition product, including: During the model application phase, the following steps are performed: acquire 1-kilometer resolution FY-4B full disk data of Shandong region at night, including data from channels C09, C10, C12, C13, C14, and C15; The FY-4B full-disk data of the Shandong region at night with a resolution of 1 km was subjected to spatiotemporal alignment, resampling and normalization processing consistent with the training phase. The processed nighttime 1km resolution Shandong region FY-4B full disk data was input into the trained U-Net model to obtain the inverted predicted C04 channel data. The C04 channel data obtained by inversion is stitched together with the 1 km resolution FY-4B full disk data of Shandong region at night. The complete channel data obtained after splicing is input into a convection initiation identification process based on a set of local temperature thresholds that are different from general thresholds and are specifically used in the Shandong region, to generate an all-weather convection initiation identification product.

Images