Meteorological satellite earth service system

Abstract

This application relates to the field of meteorological satellite technology and provides a meteorological satellite Earth operational service system, comprising: a data acquisition module for acquiring basic data; a data processing module for preprocessing the basic data; and a data product service module for publishing and providing data product services to users. The data processing module includes a data preprocessing unit, which further includes: a multi-resolution cloud image time-density encryption component for generating time-encrypted multi-resolution cloud image data; a high-resolution cloud image data reconstruction component for generating spatial high-resolution cloud image data; and a visible light cloud image generation component for generating visible light cloud image data using infrared cloud image data. This application implements functions such as encrypting low-resolution time-density cloud image data, reconstructing low-resolution cloud image data to high resolution, and automatically generating visible light cloud image data using infrared cloud image data.

Description

Technical Field

[0001] This invention relates to the field of meteorological satellite technology, and more specifically to a meteorological satellite Earth service system. Background Technology

[0002] The existing Fengyun meteorological satellite Earth operational system, by accessing data output from the ground operational systems of multiple Fengyun-3 and Fengyun-4 geostationary and polar-orbiting satellites, and relying on various product service algorithms, can provide users with five types of Earth operational data product services: cloud imagery, elements and events, severe weather, climate, and model verification. The existing Fengyun meteorological satellite Earth operational system's operational data model mainly comprises three parts: a data acquisition module, a data processing module, and a data product service module. The data acquisition module is used to acquire basic data; the data processing module is used to preprocess the basic data and generate data products from the preprocessed basic data through various algorithm units and product units; the data product service module is used to publish and provide data product services to users.

[0003] The existing Fengyun meteorological satellite Earth service system is affected by various factors, such as the varying payload performance of Fengyun meteorological satellites at different times and the inconsistency in data standards across multiple ground service systems. This results in significant differences in the parameters of the basic data acquired by the Fengyun meteorological satellite Earth service system's data acquisition module from different sources. For example, regarding cloud imagery data, the temporal and spatial resolutions of basic data from different sources differ, and some data has low temporal and spatial resolutions, failing to meet the application requirements of subsequent algorithm and product units. Furthermore, the lack of nighttime and some daytime visible light cloud images prevents the provision of full-time visible light cloud images to users. These issues collectively lead to data products generated by the data processing module sometimes being of low quality, or even failing to meet the requirements for providing data product services. Summary of the Invention

[0004] This application provides a meteorological satellite Earth service system to address the technical problems in existing Fengyun meteorological satellite Earth service systems, such as significant differences in basic cloud image parameters, low temporal and spatial resolution of some cloud image data, and the inability of cloud image data to meet the requirements for providing data product services due to the lack of nighttime and some daytime visible light cloud images.

[0005] A meteorological satellite Earth service system includes a data acquisition module 1 for acquiring basic data; a data processing module 2 for preprocessing the basic data; and a data product service module 3 for publishing and providing data product services to users. The data processing module 2 includes a data preprocessing unit 21, which further includes: a multi-resolution cloud image time encryption component 211 for generating time-encrypted multi-resolution cloud image data; a high-resolution cloud image data reconstruction component 212 for generating spatial high-resolution cloud image data; and a visible light cloud image generation component 213 for generating visible light cloud image data using infrared cloud image data.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: they can effectively solve the problems of different temporal and spatial resolutions of meteorological cloud image basic data from different sources, low temporal and spatial resolution of some data, and missing nighttime and some daytime visible light cloud images in some cloud image data. They can achieve functions such as encrypting low-resolution temporal density cloud image data, reconstructing low-resolution cloud image data with high resolution, and automatically generating visible light cloud image data using infrared cloud image data, thereby enabling the generation and provision of high-quality data product services. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is a schematic diagram of the structure of a business data model of a meteorological satellite earth service system according to an embodiment of this application;

[0009] Figure 2 This is a schematic diagram of the data processing module in an embodiment of this application;

[0010] Figure 3 This application embodiment presents a schematic diagram of a method for generating time-encrypted multi-resolution cloud map data using a multi-resolution cloud map time encryption component.

[0011] Figure 4 This application embodiment presents a schematic diagram of a method for generating spatial high-resolution cloud map data using a high-resolution cloud map data reconstruction component.

[0012] Figure 5 This application embodiment presents a schematic diagram of a method for generating visible light cloud image data using infrared cloud image data. Detailed Implementation

[0013] 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 a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to these processes, methods, products, or devices. In the following embodiments, unless otherwise specified, the "resolution" mentioned in the descriptions of multi-resolution, high-resolution, low-resolution, etc., refers to the number of pixels contained in a unit size (e.g., one inch) of the image of the cloud map (e.g., 1000*1000), and not to the frame rate in the time dimension (e.g., 10 frames / second); the "encryption" refers to increasing the "frame rate" of the cloud map, thereby reducing the time interval between two adjacent images constituting the cloud map.

[0015] like Figure 1 As shown, the existing operational data model of the Fengyun meteorological satellite Earth operation system is as follows: Figure 1As shown, the system comprises at least three parts: a data acquisition module 1, a data processing module 2, and a data product service module 3. The data acquisition module 1 is used to acquire basic data. Specifically, it acquires cloud images and other data through the existing Tianqing system via a data download unit, monitors data resources in the resource pool provided by the existing Fengyun satellite ground service system via a data arrival monitoring unit, and interacts with data through various API interfaces and WebServers interfaces via a data intervention unit. This enables the data acquisition module 1 to collect various required basic data. The data processing module 2 is used to preprocess the basic data. Specifically, it includes a data preprocessing unit 21, an algorithm unit 22, and a product unit 23. The data preprocessing unit 21 preprocesses the basic data acquired by the data acquisition module 1, such as converting it to a unified data format and providing unified data storage and retrieval functions for use by the algorithm unit 22 and the product unit 23. The algorithm unit 22 provides algorithmic support to the product unit 23. The various algorithm units 22 and product units 23 use the preprocessed basic data to generate data products. The data product service module 3 includes a product publishing unit 31, used to publish data products generated by the data processing module 2 in real time, and a data product service application unit 32, used to provide data product services and applications to users 4.

[0016] The existing Fengyun meteorological satellite Earth operational system is affected by various factors, such as the varying payload performance of Fengyun meteorological satellites at different times and the inconsistency in data standards across multiple ground operational systems. This results in significant differences in the parameters of the basic data acquired by the data acquisition module from different sources. Specifically, for cloud image data, the temporal and spatial resolutions of the basic data from different sources differ, and some data has low temporal and spatial resolutions, failing to meet the application requirements of subsequent algorithm and product units. Furthermore, the cloud image data lacks nighttime and some daytime visible light cloud images, failing to provide users with full-time visible light cloud images. These issues collectively lead to data products generated by the data processing module sometimes being of low quality, or even failing to meet the requirements for providing data product services. According to the "Work Plan for Enhancing the Application Capabilities of Fengyun Meteorological Satellites" issued by the China Meteorological Administration, the existing Fengyun meteorological satellite Earth operational system needs to be improved and perfected to address the existing problems and deficiencies.

[0017] Example 1

[0018] like Figure 1 and Figure 2 As shown, this application provides a meteorological satellite Earth service system, which includes:

[0019] Data acquisition module 1 is used to acquire basic data. Specifically, the data acquisition module 1 acquires cloud image and other data through the data download unit using the existing Tianqing system, monitors data resources in the resource pool provided by the existing Fengyun satellite ground service system through the data arrival monitoring unit, and realizes data interaction through various API interfaces, WebServers interfaces, etc. through the data intervention unit, thereby enabling the data acquisition module 1 to collect various types of required basic data.

[0020] The data processing module 2 is used to preprocess the basic data. Specifically, the data processing module 2 includes a data preprocessing unit 21, an algorithm unit 22, and a product unit 23. The data preprocessing unit 21 is used to preprocess the basic data acquired by the data acquisition module 1 for use by the algorithm unit 22 and the product unit 23. The algorithm unit 22 is used to provide algorithmic support for the product unit 23. The various algorithm units 22 and the various product units 23 use the preprocessed basic data to generate data products.

[0021] The data product service module 3 is used to publish and provide data product services to users; specifically, the data product service module 3 includes a product publishing unit 31, used to publish the data products generated by the data processing module 2 in real time, and a data product service application unit 32, used to provide data product services and applications to users 4.

[0022] The data processing module 2 includes a data preprocessing unit 21, used to preprocess the basic data acquired by the data acquisition module 1. The preprocessing includes temporal resolution encryption and spatial resolution encryption of the meteorological cloud image, and generating visible light cloud image data using infrared cloud image data. The data preprocessing unit 21 further includes:

[0023] The multi-resolution cloud image time encryption component 211 is used to generate time-encrypted multi-resolution cloud image data. By using the multi-resolution cloud image time encryption component 211, the technical problems of traditional cloud image data encryption methods, such as excessively long intervals between adjacent cloud images, rapid changes in cloud systems, unequal observation time intervals between adjacent cloud images, uneven changes in cloud systems, and abrupt transitions, are addressed, resulting in poor cloud image animation quality.

[0024] The high-resolution cloud image data reconstruction component 212 is used to generate high-resolution spatial cloud image data. The high-resolution cloud image data reconstruction component 212 solves the problem in existing technologies where the observation data resolution is too low, making it impossible to generate high-resolution cloud images. It achieves the technical effect of generating high-resolution data and cloud images, helping users to better analyze weather conditions.

[0025] The visible light cloud image generation component 213 is used to generate visible light cloud image data from infrared cloud image data. By using the visible light cloud image generation component 213, the technical problem in the prior art where the characteristic differences between daytime and nighttime infrared cloud images lead to significant differences in visible light cloud images generated at adjacent times is solved. This achieves the function of automatically generating visible light cloud images, thus improving the accuracy of visible light cloud images.

[0026] In summary, this embodiment provides a meteorological satellite Earth service system that effectively addresses issues such as varying temporal and spatial resolutions of meteorological cloud imagery from different sources, low temporal and spatial resolutions in some data, and missing nighttime and daytime visible light cloud images. It enables functions such as encrypting low-resolution temporal density cloud imagery, reconstructing low-resolution cloud imagery with high resolution, and automatically generating visible light cloud imagery from infrared cloud imagery, thereby generating and providing high-quality data product services.

[0027] Example 2

[0028] This embodiment provides a specific method for generating time-encrypted multi-resolution cloud map data using the multi-resolution cloud map time encryption component 211 described in Embodiment 1. It aims to solve the technical problems of traditional cloud map data encryption methods, such as excessively long intervals between adjacent cloud maps, rapid changes in cloud systems, unequal observation time intervals between adjacent cloud maps, uneven changes in cloud systems, and abrupt transitions, resulting in poor cloud map animation quality. Figure 3 As shown, the method includes:

[0029] A set of high-resolution cloud images is selected from a preset continuous time series, wherein the high-resolution cloud images are typical weather phenomena selected from historical data observed at high frequencies;

[0030] High-resolution cloud images containing typical weather phenomena are selected from historical high-frequency observation data. High-frequency observation data provides more detailed and accurate information. According to the preset continuous time series, a set of continuous high-resolution cloud images are selected as input. These cloud images can be selected for weather phenomena in specific regions or time periods (such as forests, cities, deserts, etc., daytime, nighttime, early morning, evening, etc., and sunny, rainy, snowy, and windy weather, etc.) and are adjacent in time or have a certain time interval to ensure their representativeness and typicality.

[0031] The high-resolution cloud map is divided into grids according to a preset image cutting size to obtain multiple high-resolution block cloud maps;

[0032] Determine the preset image cutting size according to specific needs, and divide the high-resolution cloud image into grids according to the preset cutting size. Common cutting algorithms can be used, such as cutting to the same size or cutting according to a specific ratio. Each grid corresponds to a high-resolution block cloud image. After segmentation, each high-resolution block cloud image contains part of the information of the original high-resolution cloud image, and its size conforms to the preset cutting size, resulting in multiple high-resolution block cloud images, where each block cloud image represents a part of the weather phenomena in the original high-resolution cloud image.

[0033] The high-resolution cloud map is pre-processed by shrinking to output a low-resolution cloud map;

[0034] Image downsampling methods, such as average pooling and max pooling, are used to merge or sample pixel information in the original high-resolution image to obtain a low-resolution cloud map. This reduces the detail and resolution of the image, making the cloud map smaller while retaining the main weather phenomenon features. By downsampling, a low-resolution cloud map is output. This process can reduce the complexity and storage requirements of the image, and also improve the efficiency of subsequent processing and encryption.

[0035] The low-resolution cloud map is divided into grids according to the preset image cutting size to obtain multiple low-resolution block cloud maps.

[0036] Determine the preset image cutting size, consistent with the cutting size used when previously performing grid segmentation on the high-resolution cloud image. Divide the low-resolution cloud image into grids according to the preset cutting size. The method and steps for grid segmentation are the same as when performing grid segmentation on the high-resolution cloud image. Algorithms that cut to the same size or according to a specific ratio can be used. Each grid corresponds to a low-resolution block cloud image. After segmentation, multiple low-resolution block cloud images are obtained. Each block cloud image contains some information from the original low-resolution cloud image, and its size conforms to the preset cutting size.

[0037] The multiple high-resolution block cloud maps and the multiple low-resolution block cloud maps are input into the cloud map encryption block, and the multiple high-resolution block cloud maps are encrypted by intermediate frame interpolation according to the time parameters of the preset continuous time sequence to obtain multiple encrypted high-resolution block cloud maps.

[0038] Multiple high-resolution and low-resolution block cloud maps are used as input data and fed into the cloud map encryption block for encryption processing. Based on the preset continuous time sequence parameters, intermediate frame interpolation encryption is performed on the multiple high-resolution block cloud maps. Intermediate frame interpolation refers to inserting some intermediate frames between two consecutive frames to increase the smoothness of the animation and the visual effect. Various encryption algorithms or techniques can be used in the encryption process, such as pixel mixing, motion estimation, etc. Finally, multiple encrypted high-resolution block cloud maps are obtained, which contain the intermediate frames that have undergone interpolation encryption processing.

[0039] Furthermore, the method for training the cloud map encrypted blocks includes:

[0040] Construct an encrypted training set and an encrypted test set. The encrypted training set includes high-resolution block cloud image samples and low-resolution block cloud image samples, as well as time parameters. The matrix vectors between the high-resolution block cloud image samples and the low-resolution block cloud image samples have a corresponding relationship. Based on the constructed encrypted training set, train the neural network of the frame interpolation algorithm to generate a pre-encrypted block with the goal of outputting the intermediate frame between the high-resolution block cloud image samples and the low-resolution block cloud image samples. Verify the pre-encrypted block according to the encrypted test set. When the verification is successful, output the encrypted cloud image block.

[0041] From the generated high-resolution and low-resolution patch cloud maps, a portion is selected as the encrypted training set, and another portion is selected as the encrypted test set. Both the encrypted training and test sets consist of high-resolution patch cloud map samples, low-resolution patch cloud map samples, and time parameters. The high-resolution and low-resolution patch cloud map samples have a correspondence; that is, high-resolution and low-resolution patch cloud map samples at the same location correspond to the same weather phenomenon area. The time parameters describe the time information within a preset continuous time sequence of the samples. The time parameters are taken as the proportion of the intermediate frame to the last frame time interval within the time period of any patch cloud map, expressed as:

[0042]

[0043] Among them, T s Let T be the time of the first frame within the temporal period of any block cloud map. E Let T be the time of the last frame within the temporal period of any block cloud map. m The time of the intermediate frame within the time period of any block cloud map.

[0044] Using a pre-constructed encrypted training set as input data, a neural network model with an appropriate frame interpolation algorithm is built based on a convolutional neural network. High-resolution and low-resolution block cloud image samples are input into the neural network, and the output of the neural network is set as the target for intermediate frames. The neural network model is trained using backpropagation and optimization algorithms, such as gradient descent, to generate accurate intermediate frames based on the input high-resolution and low-resolution block cloud image samples. During training, the parameters and weights of the neural network model are continuously adjusted to improve the quality and accuracy of intermediate frame generation. After training, a trained neural network model is obtained, which can be used to generate pre-encrypted blocks, which are intermediate frames generated from the input high-resolution and low-resolution block cloud image samples.

[0045] Using the encrypted test set as input, a pre-encrypted block is generated using a pre-trained neural network model with a frame interpolation algorithm. The generated pre-encrypted block is then verified by comparing and evaluating it with real samples in the test set to measure the degree of difference between the pre-encrypted block and the test set samples. If the pre-encrypted block passes the verification, meaning it matches the test set samples and meets the preset accuracy requirements, it is output as a cloud map encrypted block.

[0046] Furthermore, the generation of pre-encrypted blocks also includes:

[0047] The neural network of the frame interpolation algorithm is trained based on the constructed encrypted training set. The first intermediate frame corresponding to the high-resolution block cloud map sample and the second intermediate frame of the low-resolution block cloud map sample are obtained in each iteration round. The structural similarity between the first intermediate frame and the second intermediate frame is analyzed. Data with similarity greater than the preset similarity is identified as optimization data and used to optimize the pre-encrypted block.

[0048] Using a pre-constructed encrypted training set as training data, a neural network model based on the frame interpolation algorithm is trained. The model parameters are updated in each iteration. In each iteration, high-resolution and low-resolution block cloud image samples are input into the neural network model. Based on the output of the network model, the first and second intermediate frames of each sample are obtained. The first intermediate frame is generated from the high-resolution block cloud image sample, while the second intermediate frame is generated from the low-resolution block cloud image sample. These intermediate frames are used for model updates and optimization during training to improve the accuracy and performance of the frame interpolation algorithm.

[0049] For each pair of first and second intermediate frames, structural similarity analysis is performed. Similarity indicators include structural similarity index and mean squared error. The calculated similarity indicators are compared with a preset similarity threshold. If the similarity of a pair of intermediate frames is greater than the preset similarity threshold, the data is marked as optimized data. Optimized data is used to further improve the pre-encrypted block, for example, by adjusting parameters, retraining the model, or using other optimization methods. By processing and optimizing the optimized data, the quality and accuracy of the pre-encrypted block are improved.

[0050] The encrypted multiple high-resolution segmented cloud maps are stitched together and restored according to the preset continuous time sequence time parameters to obtain the restored encrypted high-resolution cloud map.

[0051] Furthermore, the process of stitching together and restoring the encrypted multiple high-resolution segmented cloud maps according to the preset continuous time sequence parameters includes:

[0052] Calculate the size information of the image to be stitched; create a blank image according to the size information; fill in the encrypted multiple high-resolution block cloud maps according to the time parameters of the preset continuous time sequence; and output the encrypted high-resolution cloud map.

[0053] Determine the number of block cloud maps (Q) in the image to be stitched; obtain the size information of each block cloud map, including its width and height; determine the size of the overlapping area between adjacent blocks. Based on the above information, sum the widths and heights of the Q blocks and subtract the size of the overlapping area to obtain the size information of the image to be stitched, including the width and height to be stitched.

[0054] Based on the calculated size information of the image to be stitched, a blank image is created in memory. The width of the blank image is the width to be stitched, and the height is the height to be stitched. According to the preset continuous timing parameters, multiple encrypted high-resolution block cloud images are sequentially filled into the corresponding positions in the blank image. After filling is completed, an encrypted high-resolution cloud image is obtained, which contains information of multiple encrypted block cloud images.

[0055] Furthermore, it also includes:

[0056] The multiple high-resolution block cloud maps and the multiple low-resolution block cloud maps obtained are identified by the cutting row and column numbers to generate a high-resolution identification matrix and a low-resolution identification matrix; the high-resolution identification matrix and the low-resolution identification matrix are stored in a folder named according to the data time command.

[0057] Each segmented cloud map is represented by a unique identifier using row and column numbers. The resulting multiple high-resolution and low-resolution segmented cloud maps are traversed, and each segmented cloud map is identified according to its cutting row and column numbers. High-resolution and low-resolution identifier matrices are generated respectively. The size of these matrices is the same as the number of rows and columns of the original cloud map. Each element represents the identifier of the segmented cloud map at the corresponding position. These identifier matrices are used to accurately locate and identify the segmented cloud maps, facilitating subsequent processing, storage, and restoration operations.

[0058] Create a folder named after the data time to store the high-resolution and low-resolution identification matrices. Save the high-resolution identification matrix as a file and store it in this folder. Similarly, save the low-resolution identification matrix as a file and store it in the same folder. This helps to organize and manage the relevant identification information and ensures the integrity and traceability of the data.

[0059] In summary, the encryption method and system for multi-resolution cloud images provided in this application have the following technical effects:

[0060] By selecting a set of high-resolution cloud images under a preset continuous time series and choosing representative weather phenomena as input data, the number of high-resolution cloud images is reduced while the quality of cloud images is improved.

[0061] High-resolution cloud maps are divided into grids using preset image cutting sizes and then pre-processed by shrinking to obtain low-resolution cloud maps. This reduces the dimensionality of the cloud map data and improves processing speed.

[0062] Interpolation and encryption of multiple high-resolution segmented cloud images are performed using preset continuous time parameters. This increases the temporal resolution of the cloud images, making cloud system changes smoother and more continuous.

[0063] During the restoration process, the encrypted high-resolution segmented cloud map is stitched together and restored according to the preset continuous time sequence parameters, so that the final high-resolution cloud map can be recovered, ensuring the integrity and accuracy of the data.

[0064] In summary, this encryption method solves the problems of uneven temporal resolution and insufficient temporal resolution of low-frequency observations, and achieves a smoother, more continuous, and more intuitive display of cloud system changes. This is of great significance for providing better weather forecast services and the application of cloud image animation in publicity and popular science fields.

[0065] Example 3

[0066] This embodiment provides a specific method for generating spatial high-resolution cloud image data using the high-resolution cloud image data reconstruction component 212 described in Embodiment 1, which helps users better utilize the reconstructed high-resolution cloud image data for applications such as weather analysis. Figure 4 As shown, the method includes:

[0067] Connect to the geostationary orbit radiometric imager to acquire newly added cloud image source data captured by the geostationary orbit radiometric imager;

[0068] This application utilizes a deep learning-based super-resolution reconstruction algorithm to generate high-resolution data and cloud images from low-resolution data observed by the Fengyun satellite. Fengyun satellite cloud images are widely used meteorological satellite observation products; high-resolution images allow for better observation of various details, aiding in better weather analysis and disaster forecasting, thus providing better services. For example, currently, the observation data from some channels of the Fengyun-4A satellite imager has low resolution and does not meet the requirements for generating high-resolution cloud images. Therefore, using technical means to generate high-resolution data and cloud images from low-resolution observation data helps users better analyze weather conditions.

[0069] The newly acquired cloud imagery data can be obtained by the geostationary radiometric imager. The geostationary radiometric imager is one of the main payloads of the Fengyun-4 geostationary meteorological satellite. It achieves precise and flexible two-dimensional pointing through a sophisticated dual-scan mirror mechanism, enabling rapid regional scanning within minutes. It employs an off-axis three-mirror primary optical system to acquire Earth cloud images across more than 14 spectral bands at high frequency, and utilizes an onboard blackbody for high-frequency infrared calibration to ensure the accuracy of the observation data. The newly acquired cloud imagery data refers to new satellite cloud imagery data collected by the geostationary radiometric imager. The connection is made through the geostationary radiometric imager to acquire the newly acquired cloud imagery data captured by it. Acquiring this data contributes to improving the accuracy of subsequent newly acquired cloud imagery data.

[0070] Perform block preprocessing on the newly added cloud image source data to obtain a set of cloud image blocks, where each cloud image block has a location identifier;

[0071] Block preprocessing involves dividing the newly added cloud image source data into several intervals and preprocessing each interval. Each interval refers to a segmented cloud image. The set of segmented cloud images is the collection obtained by combining the intervals. Location identifiers are the position coordinates of each segmented cloud image within the newly added cloud image source data. This block preprocessing of the newly added cloud image source data yields a set of segmented cloud images, where each segmented cloud image carries a location identifier. Obtaining this set of segmented cloud images lays the groundwork for subsequent cloud image reconstruction.

[0072] Obtain the image block unit size, where the image block unit size is 512×512;

[0073] Perform block preprocessing of the newly added cloud image source data according to the image block unit size, and mark the position of the cloud image block according to the position coordinates of the cloud image block in the newly added cloud image source data;

[0074] The cloud map block image set is generated from multiple cloud map block images with location identifiers.

[0075] The image block unit size is fixed at 512×512. Preprocessing of the newly added cloud image source data is performed according to this unit size, and the cloud image blocks are identified by their position coordinates within the new source data. A set of cloud image blocks is generated from multiple identified cloud image blocks. This set of cloud image blocks lays the groundwork for subsequent image reconstruction.

[0076] A cloud map reconstruction model is constructed based on an enhanced deep residual network;

[0077] Increasing the depth of deep networks also increases the difficulty of training. To address this increased training difficulty, a residual learning framework is proposed. This framework simplifies network training while remaining suitable for use in deeper networks. Through research and experimentation, these residual networks are easier to optimize and can maintain accuracy while increasing depth. Using the ImageNet image dataset, an error rate of only 3.57% is achieved. A cloud image reconstruction model is built based on enhanced deep residual networks, laying the groundwork for further improvements in image accuracy.

[0078] An initial cloud map reconstruction model is constructed based on an enhanced deep residual network, and the L1 loss function is used as a constraint in the supervised learning process of the initial cloud map reconstruction model.

[0079] Query the cloud image data capture records of the geostationary orbit radiation imager to obtain the historical cloud image data set;

[0080] Influencing factors were analyzed on the historical cloud map data set to obtain data influence factors;

[0081] The historical cloud map data set is filtered based on the data influence factors, and the training dataset of the cloud map reconstruction model is constructed using the filtered historical cloud map data.

[0082] The initial cloud map reconstruction model is obtained by performing supervised learning on the training dataset.

[0083] Super-resolution reconstruction aims to restore a low-resolution image sequence to its original high-resolution form. To better achieve this goal, three techniques were tested: OpenCV interpolation, the Laplacian Pyramid Regression Network (LAPSRN) based on deep learning, and Enhanced Deep Residual Network (EDSR) for single-image super-resolution. After testing these three algorithms, it was found that the deep learning-based super-resolution reconstruction algorithm significantly outperformed the OpenCV interpolation algorithm, while the Enhanced Deep Residual Network (EDSR) for single-image super-resolution slightly outperformed the Laplacian Pyramid Regression Network (LAPSRN). Therefore, the Enhanced Deep Residual Network was selected to construct the initial cloud image reconstruction model. L2 loss is the most widely used loss function for general image restoration. Experiments showed that L1 loss provided better convergence than L2 loss, while L2 loss introduced some distortion in the results. From the results, L1 loss... The loss function is slightly better than L2, so L1 loss function is used as a constraint. The historical cloud map data set refers to the cloud map data capture records of the geostationary orbit radiometric imager in the past time period. The data influence factor refers to the factors that affect the cloud map data, which in this application includes different times of the day and months. Screening refers to selecting different datasets according to different influencing factors. In the training dataset, the dataset selection target is: when selecting the dataset, data at noon should be selected as much as possible. The data selected is the visible light 500M and 2000M data of FY4A geostationary orbit radiometric imager (AGRI) NOMChannel02, which constitutes 500 pairs of sample sets, 360 pairs of training sets, and 140 pairs of test sets. The 500M data is used as HR (high resolution image) and the 2000M data is used as LR (low resolution image). Supervised learning refers to the process of adjusting parameters using a set of samples of known categories to achieve the required performance. Supervised learning is a machine learning task that infers a function from labeled training data.

[0084] An initial cloud map reconstruction model is constructed based on an enhanced deep residual network, and the L1 loss function is used as a constraint in the supervised learning process of the initial cloud map reconstruction model. A historical cloud map dataset is obtained; the influencing factors of the historical cloud map dataset are analyzed to obtain data influence factors; these data influence factors are used to filter the historical cloud map dataset, and the filtered historical cloud map data is used to construct the training dataset for the cloud map reconstruction model; supervised learning of the initial cloud map reconstruction model is performed based on the training dataset to obtain the cloud map reconstruction model. The cloud map reconstruction model trained on the training dataset has more accurate output data.

[0085] Obtain preset data influencing factors, including the data capture month and data capture period;

[0086] Based on the month of data capture, an impact analysis is performed on the historical data set of the cloud map to obtain the monthly impact deviation;

[0087] Based on the data capture period, the influence degree of the historical cloud map data set is analyzed to obtain the period influence deviation;

[0088] Set a deviation threshold, and determine the monthly impact deviation and the time period impact deviation based on the deviation threshold;

[0089] When the monthly impact deviation is greater than the deviation threshold, the month is set as the data impact factor;

[0090] When the deviation of the time period is greater than the deviation threshold, the time period is set as the data impact factor.

[0091] Preset data influencing factors refer to the factors that affect the cloud map data, set according to the impact analysis on cloud map data described above. These factors include the month and different times of day, i.e., the data capture month and data capture period. Impact degree analysis refers to the degree of influence of the data influencing factors on the cloud map data. Monthly impact bias refers to the impact of the month on the cloud map data. For example, randomly selecting 5 days of data from a certain month, approximately 500 pairs of training sets are used, and data from other months are selected according to the corresponding dates to form training sets. Twelve models are trained using the same parameters on the 12 months of data, and approximately 350 data points from the same time each day are used as the test set. The PSNR (Peak Signal-to-Noise Ratio) of the 12 models on the test set is calculated to observe the impact of different months on the models. If the PSNR difference is approximately 0.1, the impact of monthly variations can be ignored. Time period impact bias refers to the impact of different times of day on the cloud map data. For example, in the pessimistic criterion (the best algorithm for the worst time), the training performance of the models at different times of day is compared, and the UTC time is found to be the best time. Data around 04:30 is best suited for the training set (due to ample midday sunlight in China, cloud image details are richest); the deviation threshold is set by staff based on experience and is used to analyze the impact of the month and time period; the data impact factors are analyzed by comparing the impact of the month and time period with the deviation threshold.

[0092] Preset data influencing factors are obtained, including the data capture month and data capture period. The influence of the data capture month on the historical cloud map data set is analyzed to obtain the monthly influence deviation. Similarly, the influence of the data capture period on the historical cloud map data set is analyzed to obtain the period influence deviation. A deviation threshold is set, and the monthly and period influence deviations are judged based on this threshold. When the monthly influence deviation is greater than the deviation threshold, the month is set as the data influence factor; when the period influence deviation is greater than the deviation threshold, the period is set as the data influence factor. The data influence factors obtained through analysis and comparison lay the foundation for subsequent model training.

[0093] like Figure 2 As shown, data preprocessing of the training dataset is performed, including data cleaning, format conversion, image segmentation, and edge smoothing.

[0094] The preprocessed training dataset is then augmented to obtain an augmented training dataset.

[0095] Using the L1 loss function as a constraint during model training, and based on the ADAM optimizer, the model is trained using the augmented training dataset to generate the cloud map reconstruction model.

[0096] Data preprocessing refers to transforming and processing data using various methods to make it suitable for storage, management, and further analysis and application; data cleaning is the final procedure in data files to find and correct identifiable errors, including checking data consistency and handling invalid and missing values; format conversion is the process of converting data from one format or structure to another; image segmentation is the process of dividing an image into several semantic targets. When the resolution of the image to be processed is too large and resources are limited (such as video memory, computing power, etc.), the image can be divided into small blocks; edge smoothing refers to fitting the original data to generate smoother data, generally to make the data more readable and interpretable; augmentation refers to simulating sampling of the training dataset according to an ideal data distribution. Before training, data cleaning (removing padding / invalid values ​​and normalization) is performed, and training is conducted separately for different satellite types. HDF data is used directly instead of JPG images as samples during training to reduce feature loss. During training, 32×32 input blocks from low-resolution images and corresponding 128*128 input blocks from high-resolution images are used, and the images are randomly flipped and rotated 90° around their center points to enhance the training data.

[0097] In the training of the cloud image reconstruction model, the batch size was set to 4, the learning rate was initialized to 1e-4, and the ADAM optimizer was used to train the model. Training was performed on an NVIDIA Tesla M60 8G GPU. Due to server performance limitations, 100 epochs required 3 days, and some parameters were subject to certain restrictions. During the process of restoring low-resolution data to high-resolution data, server performance limitations prevented direct super-resolution of the complete data. The data needed to be split into smaller chunks, processed, and then the results stitched together to generate the required augmented training dataset. The model was trained using this augmented training dataset, generating the cloud image reconstruction model, which facilitates subsequent reconstruction of satellite cloud images.

[0098] The cloud image reconstruction model sequentially performs image reconstruction of the cloud image block set to generate an optimized cloud image block set.

[0099] Image reconstruction refers to stitching together the cloud map blocks sequentially according to the order of each block. This can be one or more sequences, such as dividing the image into multiple sub-regions or stitching them together at multiple levels. The cloud map reconstruction model sequentially performs image reconstruction on the set of cloud map blocks to generate an optimized set of cloud map blocks. Generating an optimized set of cloud map blocks lays the groundwork for generating subsequent cloud map reconstruction data.

[0100] Based on the location identifier, the block image stitching of the optimized cloud map block image set is performed to generate cloud map reconstruction data.

[0101] Based on the location identifiers of the optimized cloud image patch set, the patch images of the optimized cloud image patch set are stitched together to obtain cloud image reconstruction data. Cloud image reconstruction data refers to the use of technical means to generate high-resolution data and cloud images from observed low-resolution data, helping users to better analyze weather conditions.

[0102] Construct a segmented image stitching sequence, wherein the segmented image stitching sequence is one or more;

[0103] Based on the location identifier, the block images in the optimized cloud map block image set are arranged according to the block image stitching sequence to obtain the optimized cloud map block image stitching sequence;

[0104] Based on the optimized cloud map segmented image stitching sequence, segmented image stitching is performed to generate the cloud map reconstruction data.

[0105] The images can be stitched together sequentially according to their order, and this can be one or more sequences, such as dividing the image into multiple sub-regions or multiple layers. Based on the location identifier, the images in the optimized cloud map image set are arranged according to the image stitching sequence to obtain the optimized cloud map image stitching sequence. Image stitching is then performed according to the optimized cloud map image stitching sequence to generate the cloud map reconstruction data. This application solves the problem of low-resolution observation data in the prior art, which prevents the generation of high-resolution cloud maps, and achieves the technical effect of generating high-resolution data and cloud maps, helping users to better analyze weather conditions.

[0106] Example 4

[0107] This embodiment provides a specific method for the visible light cloud image generation component 213 described in Embodiment 1 to generate visible light cloud image data using infrared cloud image data. The aim is to solve the technical problem in the prior art where the characteristic differences between daytime and nighttime infrared cloud images lead to significant differences in visible light cloud images generated at adjacent times. For example... Figure 5 As shown, the method includes:

[0108] S1: Satellite observation is performed based on a multi-channel scanning imaging radiometer to collect a nighttime infrared cloud image set of targets to be converted into cloud images. The nighttime infrared cloud image set of targets corresponds to different channel bands.

[0109] In one embodiment of this application, the multi-channel scanning imaging radiometer is used to collect multiple spectral radiation information of the Earth's surface and atmosphere, providing data for weather forecasting and climate monitoring. It features a large number of detection bands, spatial resolution, and temporal resolution, enabling multispectral, high-frequency, and quantitative detection of physical parameters of the Earth's surface and atmosphere. Although the infrared channel in the multi-channel scanning imaging radiometer can perform observations around the clock, the infrared cloud images from daytime and nighttime show significant differences in the characteristics of different underlying surfaces. Therefore, it is necessary to convert the nighttime target infrared cloud image atlas to eliminate the unique characteristic differences between daytime and nighttime infrared data.

[0110] Preferably, the multi-channel scanning imaging radiometer has 14 channels, including 6 visible / near-infrared bands, 2 mid-infrared bands, 2 water vapor bands, and 4 long-infrared bands. The nighttime target infrared cloud image set to be converted is data acquired by the infrared channels of the multi-channel scanning imaging radiometer. Each channel is two-dimensional grid data, and each grid point stores a positive integer value between 0 and 4095. The nighttime target infrared cloud image set corresponds to different channel bands. By acquiring the nighttime target infrared cloud image set to be converted, the goal of providing data for subsequent visible light cloud image generation is achieved.

[0111] S2: Based on the nighttime target infrared cloud image set, divide the cloud image data into blocks and obtain the block infrared cloud image data;

[0112] Furthermore, such as Figure 2 As shown, the step S2 of this embodiment of the application, which involves dividing the cloud image data into blocks and obtaining the block infrared cloud image data, includes:

[0113] S2-1: Establish a two-dimensional coordinate system using the cloud map coverage area as the coordinate space;

[0114] S2-2: Based on the predetermined segmentation size, and combined with the two-dimensional coordinate system, the nighttime infrared cloud image set is traversed to perform uniform segmentation of the cloud image, and multi-group block cloud images are obtained, wherein the multi-group block cloud images have neighboring overlapping areas.

[0115] S2-3: The multi-group block cloud map is used as the block infrared cloud map data, and the multi-group block cloud map corresponds one-to-one with the nighttime target infrared cloud map set.

[0116] In one possible embodiment, the images in the nighttime target infrared cloud image set are divided into cloud image data blocks to lay the groundwork for improving the efficiency and accuracy of subsequent cloud image conversion. The block infrared cloud image data is obtained by dividing the nighttime target infrared cloud image set into blocks according to certain segmentation rules.

[0117] Using the cloud image coverage area as the coordinate space, a two-dimensional coordinate system is constructed to describe the image positions within the cloud image coverage area. The cloud image coverage area refers to the region covered by the acquired nighttime target infrared cloud image set. The predetermined segmentation size is a segmentation scale determined by those skilled in the art for uniformly segmenting the cloud image, and can be a size of 1024×1024. Using the two-dimensional coordinate system as the data basis for division, the nighttime infrared cloud images in each nighttime infrared cloud image set are uniformly divided according to the predetermined segmentation size to obtain corresponding block results, i.e., the multi-group block cloud images. Each group of block cloud images corresponds to one nighttime infrared cloud image set. Preferably, during the uniform segmentation of the cloud image, adjacent two block cloud images have overlapping neighborhood areas, which facilitates the fusion of the block infrared cloud image data and reduces the generation of seams. Furthermore, the multi-group block cloud images are used as the block infrared cloud image data to provide data for subsequent block stitching.

[0118] S3: Call the seasonal time-domain cloud map within the predetermined time interval, the seasonal time-domain cloud map including the nighttime infrared cloud map, the daytime infrared cloud map and the daytime visible light cloud map;

[0119] In one embodiment, to obtain data for training a differential adaptive calibration model that performs differential analysis on nighttime and daytime cloud images, the seasonal temporal cloud image is obtained by calling images observed by a multi-channel scanning imaging radiometer within a predetermined time interval. Preferably, the predetermined time interval can be hourly cloud images within 10 days before and after the spring equinox, summer solstice, autumn equinox, and winter solstice, thus covering both day and night. The seasonal temporal cloud image within the predetermined time interval includes nighttime infrared cloud images, daytime infrared cloud images, and daytime visible light cloud images. Specifically, the nighttime infrared cloud image is acquired at night using the infrared channel of the multi-channel scanning imaging radiometer. The daytime infrared cloud image is acquired during the day using the infrared channel of the multi-channel scanning imaging radiometer. The daytime visible light cloud image is acquired during the day using the visible light channel of the multi-channel scanning imaging radiometer.

[0120] S4: Based on the nighttime infrared cloud image and the daytime infrared cloud image, train a differential adaptive calibration model for differential analysis and processing of nighttime and daytime cloud images;

[0121] Furthermore, such as Figure 3As shown, the differential adaptive calibration model for training nighttime and daytime cloud image differential analysis, in this embodiment of the application, step S4 includes:

[0122] S4-1: Based on Siamese networks, supervised training enables nighttime and daytime infrared feature processing channels that can perform cloud map translation and convolutional feature extraction.

[0123] S4-2: Perform parallel channel deployment and apply a differential compensation layer to generate the adaptive calibration model;

[0124] S4-3: Based on the daytime infrared feature processing channel, perform mapping and matching of the input infrared cloud image of the nighttime infrared feature processing channel, and simultaneously perform dual-channel cloud image processing.

[0125] Furthermore, step S4 in this embodiment of the application also includes:

[0126] S4-4: Based on the seasonal time-domain cloud map, extract the nighttime infrared cloud map and the daytime infrared cloud map, perform analysis and processing to obtain the calibrated converted cloud map, and map to determine the training data;

[0127] S4-5: Based on the training data, train and generate an initial adaptive calibration model;

[0128] S4-6: Based on the training data, the initial adaptive calibration model is tested, and training data that does not meet the deviation threshold is selected for retraining until all test results meet the deviation threshold.

[0129] In one possible embodiment, by using the nighttime infrared cloud image and the daytime infrared cloud image as training data, a differential adaptive calibration model for differential analysis processing of the nighttime and daytime cloud images is trained under supervision, thereby achieving the technical effect of intelligently analyzing the differences between daytime and nighttime cloud images and improving analysis efficiency and accuracy.

[0130] Preferably, supervised training is performed using a Siamese network to train a nighttime infrared feature processing channel for cloud image translation and convolutional feature extraction of nighttime infrared cloud images, and a daytime infrared feature processing channel for cloud image translation and convolutional feature extraction of daytime infrared cloud images. The nighttime and daytime infrared feature processing channels are deployed in parallel, and the output layers of both channels are communicatively connected to the input layer of the differentiation compensation layer, thereby generating the adaptive calibration model.

[0131] Optionally, the nighttime infrared cloud image is input into the nighttime infrared feature processing channel for data processing. Simultaneously, the features extracted from the nighttime infrared cloud image are mapped and matched with the daytime infrared cloud image in the daytime infrared feature processing channel. The successfully matched daytime infrared cloud image is then used to perform pixel-level compensation on the nighttime infrared cloud image using differential compensation features. This compensates for the differences between the nighttime and daytime infrared cloud images caused by the different underlying surfaces, thus performing dual-channel cloud image processing. This achieves the technical effect of improving the reliability of the analyzed data.

[0132] Preferably, the nighttime infrared cloud image and the daytime infrared cloud image are extracted based on the seasonal time-domain cloud image, analyzed and processed to obtain a calibrated converted cloud image, and a mapping relationship is constructed between the calibrated converted cloud image and the nighttime and daytime infrared cloud images according to the correspondence, thereby determining the training data. Optionally, the training data is divided into a training set and a validation set according to a certain ratio. The training set accounts for 30%, and the validation set accounts for 70%. Supervised training is performed using the training set to obtain the initial adaptive calibration model, and then the initial adaptive calibration model is tested using the validation set. Preferably, the current model parameters are saved after each training cycle, overwriting the previous model parameters, and saved as the latest model. During the validation process, training data that does not meet the deviation threshold are screened for retraining until the test results all meet the deviation threshold, at which point the validation is passed and the differential adaptive calibration model is obtained.

[0133] S5: Based on the daytime infrared cloud image and the daytime visible light cloud image, train a cloud image conversion model to simulate the conversion between infrared cloud image and visible light cloud image;

[0134] In one embodiment, the daytime infrared cloud image and the daytime visible light cloud image are used as the first training data to supervise the training of a conditional generative adversarial network (GAN) model. The GAN model can be pix2pixHD. After training a certain number of times using the training set in the first training data, simulated visible light data is generated using all infrared data from the validation set in the first training data. PSNR is calculated using this simulated visible light data and all visible light data from the validation set in the first training data. At the end of each training cycle, the current model parameters are saved, overwriting the previous model parameters; this is the latest model. During each validation, the model with a higher PSNR than the previous one is considered the best, and the parameters of the best model are saved. The cloud image conversion model is used to convert the daytime infrared cloud image into a daytime visible light cloud image.

[0135] S6: Combining the difference adaptive calibration model with the cloud image conversion model, the segmented infrared cloud image data is processed independently to obtain the segmented converted cloud image;

[0136] Furthermore, step S6 in this embodiment of the application also includes:

[0137] S6-1: Perform image translation on the nighttime infrared cloud image and the daytime infrared cloud image to extract semantic conversion information;

[0138] S6-2: Perform differential verification of semantic conversion information and obtain the pixel domain of cloud map deviation;

[0139] S6-3: Locate the deviation pixel domain of the daytime infrared cloud image and extract convolutional features, replace the convolutional features of the deviation pixel domain in the nighttime infrared cloud image, and obtain the block-converted cloud image.

[0140] In one possible embodiment, the difference adaptive calibration model and the cloud image conversion model are connected sequentially. Then, the segmented infrared cloud image data is input into the difference adaptive calibration model and the cloud image conversion model in sequence to realize independent processing and conversion of the segmented infrared cloud image data, thereby obtaining the converted segmented converted cloud image.

[0141] In one embodiment, a deep learning algorithm for image translation is used to translate nighttime and daytime infrared cloud images to obtain semantic conversion information. Then, the semantic conversion information corresponding to the nighttime and daytime infrared cloud images is differentially verified to obtain a cloud image deviation pixel domain. This cloud image deviation pixel domain refers to the pixel region with a large deviation.

[0142] By locating the daytime infrared cloud image based on the deviation pixel domain, determining the deviation pixel domain, and extracting convolutional features, the extracted convolutional features are used to replace the nighttime infrared cloud image, thereby obtaining a segmented transformed cloud image after compensating for the differences.

[0143] S7: The segmented cloud images are stitched together to generate a daytime target visible light cloud image set.

[0144] Furthermore, in the step S7 of this embodiment, stitching together the segmented transformation cloud map further includes:

[0145] S7-1: In the two-dimensional coordinate system, the obtained block transformation cloud map is located and filled in to obtain the spliced ​​cloud map;

[0146] S7-2: Identify the stitched cloud map and extract the overlapping areas;

[0147] S7-3: Perform a weighted average fusion of the overlapping cloud map and the second overlapping cloud map in the neighboring overlapping areas to determine the regional fusion result.

[0148] In one embodiment, the obtained segmented transformed cloud images are stitched together to obtain the daytime target visible light cloud atlas. Weighted average fusion of overlapping areas of adjacent segments is used to avoid seams between adjacent segments.

[0149] Specifically, the obtained segmented transformation cloud map is located and filled in according to the corresponding points in the two-dimensional coordinate system to obtain a stitched cloud map. Based on the stitched cloud map, the overlapping areas of two adjacent segmented transformation cloud maps are extracted. The region fusion result is determined by using an inverse distance linear weighted average to fuse the overlapping cloud map of the neighboring overlapping area with the second overlapping cloud map. Preferably, for example, when segment A and segment B overlap, the value of any grid point after overlap is calculated using P = a*(1-d / L) + b*d / L, where a and b are the values ​​of segment A and segment B at that grid point, respectively, L is the width of the overlapping area, d is the distance of the grid point from the edge of the overlapping area on the a side, and P is the value of any grid point in the fused overlapping area.

[0150] Furthermore, step S7 in this embodiment of the application also includes:

[0151] S7-4: Measure and calculate the solar altitude angle, and determine the time point for cloud image acquisition;

[0152] S7-5: Determine the cloud map acquisition mode based on the cloud map acquisition time point;

[0153] S7-6: If it is daytime, directly acquire visible light satellite cloud images as the first cloud image acquisition mode;

[0154] S7-7: If it is nighttime, collect infrared satellite cloud images and perform compensation conversion to obtain visible light satellite cloud images as the second cloud image acquisition mode;

[0155] S7-8: If it is the neighborhood of the terminator, execute the first cloud image acquisition mode and the second cloud image acquisition mode, and perform a weighted average of the execution results to obtain a visible light satellite cloud image.

[0156] In one possible embodiment, the solar altitude angle refers to the angle between the incident direction of sunlight and the horizontal plane at a certain location on Earth. The solar altitude angle varies with local time (hour angle) and solar declination; therefore, the time point for cloud image acquisition can be determined by determining the solar altitude angle. Specifically, the cloud image acquisition time point is the time point at which the multi-channel scanning imaging radiometer acquires cloud images. Preferably, the noon solar altitude angle = 90° - the latitude difference between the location and the subsolar point.

[0157] Furthermore, the mode for acquiring satellite cloud images can be determined based on the cloud image acquisition time point. When the cloud image acquisition time point is daytime, that is, when the solar altitude angle is greater than the threshold (e.g., 0 degrees), visible light satellite cloud images can be directly acquired. The acquired satellite cloud image at this time is the visible light satellite cloud image, which is used as the first cloud image acquisition mode. When the cloud image acquisition time point is nighttime, that is, when the solar altitude angle is less than the threshold (e.g., 0 degrees), the acquired nighttime infrared cloud image cannot be directly used for visible light conversion. The acquired infrared satellite cloud image needs to be compensated and converted to obtain the visible light satellite cloud image, which is used as the second cloud image acquisition mode. Preferably, when the solar altitude angle is within a specified threshold range (e.g., 0 degrees to 5 degrees), it is considered to be the terminator neighborhood. By simultaneously executing the first and second cloud image acquisition modes and weighted averaging the execution results, the result is used as the visible light satellite cloud image. For example, a weighted average is calculated based on the formulas w = a / 5 and V = w*T + (1-w)*F, where a is the solar altitude angle, w is the weight, T is the value of the actual visible light (the value obtained by executing the first cloud map acquisition mode), F is the value of the simulated visible light (the value obtained by executing the second cloud map acquisition mode), and V is the visible light value after weighted average.

[0158] In summary, the embodiments of this application have at least the following technical effects:

[0159] This application divides the nighttime target infrared cloud image atlas into data blocks, processes and transforms each block independently, and then utilizes a differentiated adaptive calibration model and cloud image transformation model based on seasonal temporal cloud images to provide intelligent processing and transformation. Finally, the transformed cloud images are stitched together to obtain a daytime target visible light cloud image atlas. This achieves the technical effect of improving the reliability of the visible light cloud image atlas, increasing generation efficiency, and enhancing its intelligence.

[0160] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, specific embodiments have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims can be performed in a different order than that shown in the embodiments and still achieve the desired result. Additionally, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0161] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0162] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A meteorological satellite Earth service system, characterized in that, include: The data acquisition module is used to obtain basic data; The data processing module is used to preprocess the basic data; as well as, The data product service module is used to publish and provide data product services to users; The data processing module includes a data preprocessing unit, which further includes: A multi-resolution cloud map time-encryption component is used to generate time-encrypted multi-resolution cloud map data; High-resolution cloud map data reconstruction component, used to generate spatial high-resolution cloud map data; Visible light cloud image generation component, used to generate visible light cloud image data from infrared cloud image data; The method for generating time-encrypted multi-resolution cloud map data by the multi-resolution cloud map time encryption component includes the following steps: A set of high-resolution cloud images is selected from a preset continuous time series, wherein the high-resolution cloud images are typical weather phenomena selected from historical data observed at high frequencies; The high-resolution cloud map is divided into grids according to a preset image cutting size to obtain multiple high-resolution block cloud maps; The high-resolution cloud map is pre-processed by shrinking to output a low-resolution cloud map; The low-resolution cloud map is divided into grids according to the preset image cutting size to obtain multiple low-resolution block cloud maps. The multiple high-resolution block cloud maps and the multiple low-resolution block cloud maps are input into the cloud map encryption block, and the multiple high-resolution block cloud maps are encrypted by intermediate frame interpolation according to the time parameters of the preset continuous time sequence to obtain multiple encrypted high-resolution block cloud maps. The encrypted multiple high-resolution block cloud maps are stitched together and restored according to the preset continuous time sequence time parameters to obtain the restored encrypted high-resolution cloud map. The method for generating spatial high-resolution cloud map data using the high-resolution cloud map data reconstruction component includes the following steps: Acquire new cloud imagery data captured by the geostationary orbit radiometric imager; Perform block preprocessing on the newly added cloud image source data to obtain a set of cloud image blocks, where each cloud image block has a location identifier; A cloud map reconstruction model is constructed based on an enhanced deep residual network; The cloud image reconstruction model sequentially performs image reconstruction of the cloud image block set to generate an optimized cloud image block set. Based on the location identifier, perform block image stitching of the optimized cloud map block image set to generate cloud map reconstruction data; The method for generating visible light cloud image data using infrared cloud image data by the visible light cloud image generation component includes the following steps: Satellite observations are performed using a multi-channel scanning imaging radiometer to collect nighttime infrared cloud images of targets to be converted into cloud images. These nighttime infrared cloud images correspond to different channel bands. Based on the nighttime target infrared cloud image set, the cloud image data is divided into blocks to obtain the block infrared cloud image data; Call up seasonal time-domain cloud maps within a predetermined time interval, the seasonal time-domain cloud maps including nighttime infrared cloud maps, daytime infrared cloud maps and daytime visible light cloud maps; Based on the nighttime infrared cloud image and the daytime infrared cloud image, a differential adaptive calibration model is trained to perform differential analysis and processing of nighttime and daytime cloud images; Based on the daytime infrared cloud image and the daytime visible light cloud image, train a cloud image conversion model to simulate the conversion between infrared cloud image and visible light cloud image; By combining the difference adaptive calibration model and the cloud image conversion model with the front and rear position connections, the segmented infrared cloud image data is processed independently to obtain the segmented converted cloud image; The segmented cloud images are stitched together to generate a daytime target visible light cloud image set.

2. The meteorological satellite Earth service system according to claim 1, characterized in that, The method for training the cloud map encrypted blocks includes: Construct an encrypted training set and an encrypted test set. The encrypted training set includes high-resolution block cloud map samples and low-resolution block cloud map samples, as well as time parameters. The matrix vectors between the high-resolution block cloud map samples and the low-resolution block cloud map samples have a corresponding relationship. The neural network of the frame interpolation algorithm is trained based on the constructed encryption training set, with the goal of outputting the intermediate frames of the high-resolution block cloud map samples and the low-resolution block cloud map samples, to generate pre-encrypted blocks. The pre-encrypted block is verified according to the encryption test set. Once the verification is successful, the cloud map encrypted block is output.

3. The meteorological satellite Earth service system according to claim 2, characterized in that, The time parameter is taken as the proportion of the intermediate frame in the time interval of the last frame within the time period of any block cloud map, and the expression is: ； in, The time of the first frame within the temporal period of any block cloud map. Let be the time of the last frame within the time period of any block cloud map. The time of the intermediate frame within the time period of any block cloud map.

4. The meteorological satellite Earth service system according to claim 3, characterized in that, The method for generating pre-encrypted blocks further includes: The neural network of the frame interpolation algorithm is trained based on the constructed encrypted training set to obtain the first intermediate frame corresponding to the high-resolution block cloud map sample and the second intermediate frame of the low-resolution block cloud map sample in each iteration round. The structural similarity between the first intermediate frame and the second intermediate frame is analyzed, and data with a similarity greater than a preset similarity are identified as optimization data and used to optimize the pre-encrypted block.

5. The meteorological satellite Earth service system according to claim 1, characterized in that, The step of performing block preprocessing of the newly added cloud map source data also includes: Obtain the image block unit size, where the image block unit size is 512×512; Perform block preprocessing of the newly added cloud image source data according to the image block unit size, and mark the position of the cloud image block according to the position coordinates of the cloud image block in the newly added cloud image source data; The cloud map block image set is generated from multiple cloud map block images with location identifiers.

6. The meteorological satellite Earth service system according to claim 5, characterized in that, The cloud map reconstruction model based on the enhanced deep residual network also includes: An initial cloud map reconstruction model is constructed based on an enhanced deep residual network, and the L1 loss function is used as a constraint in the supervised learning process of the initial cloud map reconstruction model. Query the cloud image data capture records of the geostationary orbit radiation imager to obtain the historical cloud image data set; Influencing factors are analyzed by examining the historical cloud map data set to obtain data influence factors; The historical cloud map data set is filtered based on the data influence factors, and the training dataset of the cloud map reconstruction model is constructed using the filtered historical cloud map data. The initial cloud map reconstruction model is obtained by performing supervised learning on the training dataset.

7. The meteorological satellite Earth service system according to claim 1, characterized in that, The method for segmenting cloud image data and obtaining segmented infrared cloud image data includes: A two-dimensional coordinate system is constructed using the cloud map coverage area as the coordinate space; Based on a predetermined segmentation size, and in conjunction with the two-dimensional coordinate system, the nighttime infrared cloud atlas is traversed to uniformly segment the cloud image, thereby obtaining multi-group block cloud images, which have neighboring overlapping regions. The multi-group block cloud image is used as the block infrared cloud image data, and the multi-group block cloud image corresponds one-to-one with the nighttime target infrared cloud image set.

Images