Cloud thickness identification method and system based on combination of Cycle-GAN and GNN

Abstract

The invention discloses a cloud thickness recognition method and system based on combination of Cycle-GAN and GNN, and relates to the field of computer vision and pattern recognition, and the method comprises the steps: collecting meteorological data of a target region, and carrying out the preprocessing and feature extraction; using Cycle-GAN to convert the extracted features, using a graph neural network to analyze the converted data, and extracting the space and time relationship of the cloud data; and predicting the cloud thickness by using the GNN model, and generating a prediction result of the cloud thickness of the target area. According to the invention, through multi-device collaborative observation and comprehensive utilization of multi-source data, the comprehensiveness and time-space synchronism of cloud data are improved; through multi-dimensional feature extraction, the description capability of the cloud is enhanced, and richer information is provided for cloud thickness identification; by means of a joint model architecture, data conversion and graph structure analysis are optimized, the recognition precision and reliability are remarkably improved, and through a continuous optimization mechanism, good performance of the model in a continuously changing environment is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of computer vision and pattern recognition, specifically to a cloud thickness recognition method and system based on the joint use of Cycle-GAN and GNN. Background Technology

[0002] With the rapid development of machine vision and artificial intelligence technologies, image processing and analysis have been widely applied in various fields, especially in intelligent sensing, remote sensing image processing, and environmental monitoring. Cloud thickness, as a key parameter in meteorological analysis, is of great significance for aerospace, climate prediction, and smart agriculture. Traditional cloud thickness identification methods rely on single types of observation equipment, such as meteorological satellites or ground-based radar, and employ simple image processing algorithms, such as edge detection or threshold-based segmentation methods, which cannot fully extract the complex texture, geometric, and spectral features in cloud images.

[0003] Existing technologies suffer from problems such as inconsistent image formats and spatiotemporal resolution when processing multi-source heterogeneous meteorological image data, resulting in insufficient accuracy and robustness in cloud thickness identification. Furthermore, traditional image processing-based identification methods often neglect the spatial and temporal relationships reflected in the images, making it difficult to adapt to the dynamic changes of clouds under different environmental conditions.

[0004] In recent years, Generative Adversarial Networks (GANs) and Graph Neural Networks (GNNs) in deep learning have demonstrated powerful capabilities in image style transfer, feature extraction, and relation modeling. However, existing applications rarely combine Cycle-GANs and GNNs for processing meteorological image data, especially in the area of ​​high-precision cloud thickness prediction based on image visual data, which remains a gap.

[0005] Therefore, there is an urgent need for a technical solution that combines machine vision and deep learning to efficiently process multi-source meteorological image data, improve the analysis accuracy of cloud images, and especially enhance the ability to intelligently identify cloud thickness in complex environments. Summary of the Invention

[0006] In view of the above-mentioned problems, the present invention is proposed.

[0007] Therefore, the technical problem solved by this invention is: how to solve the problems of inconsistent formats and spatial-temporal alignment of multi-source heterogeneous meteorological data by combining Cycle-GAN and GNN models, so as to achieve high-precision cloud thickness identification and prediction.

[0008] To address the aforementioned technical problems, this invention provides the following technical solution: a cloud thickness recognition method based on a combination of Cycle-GAN and GNN, comprising the following steps:

[0009] Use sensors to collect meteorological data of the target area and perform preprocessing;

[0010] Feature extraction based on preprocessed data;

[0011] Cycle-GAN is used to transform the extracted features, eliminating format differences between data sources;

[0012] Graph neural networks are used to analyze the transformed data and extract the spatial and temporal relationships of cloud data.

[0013] The GNN model is used to predict cloud thickness and generate prediction results for cloud thickness in the target area.

[0014] The use of sensors to collect meteorological data of the target area includes...

[0015] Using meteorological satellites, ground-based radar, and lidar, temperature data is continuously collected from the target area according to the preset observation plan and parameter settings.

[0016] Meteorological data, including humidity and air pressure, are collected for the target area.

[0017] The preprocessing involves unifying and calibrating the format of the collected multi-source heterogeneous data, standardizing the data according to the characteristics of data collected by different devices, mapping the data to a unified numerical range and coordinate system, and identifying the data formats of different data sources and converting them to NetCDF format.

[0018] The feature extraction based on preprocessed data includes cloud texture features, spectral features, geometric features, and features related to atmospheric environmental parameters.

[0019] As a preferred embodiment of the cloud thickness recognition method based on Cycle-GAN and GNN described in this invention, the cloud texture features are extracted using a gray-level co-occurrence matrix, and the key parameters for calculating the gray-level co-occurrence matrix are determined. The displacement vector is represented as follows:

[0020]

[0021] Where, d x d y These are the spatial distance and orientation between pixel pairs, respectively;

[0022] By determining the number of gray levels n, the gray range of the image is quantized into n levels, reducing the amount of computation and highlighting texture features. The image is quantized, and the gray value of each pixel is remapped to a small integer range according to the number of gray levels n.

[0023] The spectral features are calculated by taking the ratio of radiant intensity in each band and the difference in radiant intensity in each band, and then performing a combined feature calculation.

[0024] The geometric features are extracted by creating the shape of the clouds, obtaining the cloud contours using the Canny operator edge detection algorithm, detecting edges based on image grayscale changes, applying Gaussian filtering to the cloud image, and calculating the gradient magnitude G in the horizontal and vertical directions. m The edge pixels are determined by non-maximum suppression and double thresholding based on the gradient direction θ, thus obtaining the outline of the cloud.

[0025] The area features of clouds are extracted, the cloud images are binarized using the pixel counting method, the number of cloud pixels in the image is counted, the temperature information of the cloud top is obtained using satellite remote sensing, and the height of the cloud is estimated by combining the atmospheric temperature vertical profile model.

[0026] The features related to the atmospheric environmental parameters are the difference between cloud top temperature and ambient temperature, and the influence of humidity on cloud boundaries.

[0027] As a preferred embodiment of the cloud thickness recognition method based on Cycle-GAN and GNN as described in this invention, the step of using Cycle-GAN to transform the extracted features involves forming a feature sample set based on the collected data and the extracted features, and then performing normalization processing.

[0028] If the data do not match perfectly in space and time, alignment operations are performed. When satellite data and radar data have different spatial resolutions, they are adjusted to the same resolution using interpolation methods, integrating data from different data sources into a comprehensive dataset.

[0029] As a preferred embodiment of the cloud thickness recognition method based on the joint Cycle-GAN and GNN described in this invention, the step of using Cycle-GAN to transform the extracted features further includes constructing a Cycle-GAN generator, specifically three generators G... 12 Converting satellite data and atmospheric environmental parameters into a form similar to radar data, G 23 Converting radar data and atmospheric environmental parameters into a format similar to lidar data, G 31 The generator uses a multi-layer convolutional neural network structure to convert lidar data and atmospheric environmental parameters into a form similar to satellite data.

[0030] Construct three discriminators D1, D2, and D3 3, These are used to distinguish between genuine and fake satellite data, radar data, and lidar data, respectively.

[0031] The generator's total loss is a weighted sum of the adversarial loss and the cycle consistency loss, expressed as:

[0032]

[0033] in, The generator's total loss is a weighted sum of the adversarial loss and the cycle consistency loss. Combating losses Cyclic consistency loss, D norm Integrate data from different data sources into a single comprehensive dataset. For normalized data (D) norm The probability distribution function of p, where r represents the real data sample. data (r) represents the probability distribution function of the real data (r), and discriminators D1, D2, and D3 are used to distinguish between real and fake satellite data, radar data, and lidar data, respectively. λ cyc It is the weight of the cycle consistency loss, and the loss of the discriminator is the adversarial loss;

[0034] The training process uses the Adam optimizer to train the generator and discriminator alternately. First, the generator is fixed and the discriminator is updated. Then, the discriminator is fixed and the generator is updated. The training is iterated for a certain number of rounds until the model converges.

[0035] Using data points transformed by Cycle-GAN as nodes, edges are established based on spatial distance and data correlation, and a graph convolutional network is selected.

[0036] Each node's feature vector contains its corresponding data value and its spatial location information;

[0037] An edge is established if the spatial distance between two nodes is less than a threshold, or the similarity of their feature vectors is greater than a threshold.

[0038] Define the output layer and loss function. Add a linear layer to the last layer of the GNN to output the cloud thickness prediction. The loss function uses mean squared error, and its expression is:

[0039]

[0040] Among them, y i This is the actual cloud thickness value. This refers to the predicted cloud thickness value, where N is the number of samples. An alternating training approach is used: first, the Cycle-GAN is trained for several rounds; then, the parameters of the Cycle-GAN are fixed to train the GNN. Based on the performance of the GNN on the validation set, the weights of the generator and discriminator of the Cycle-GAN are adjusted, and the GNN is trained again in this repeated process until the joint model meets the requirements on the test set.

[0041] The performance of the joint model is evaluated using a test set, and evaluation metrics are calculated. Based on the results, the hyperparameters of the GNN cloud thickness prediction model are adjusted to improve model performance.

[0042] As a preferred embodiment of the cloud thickness identification method based on Cycle-GAN and GNN as described in this invention, the method of using the GNN model to predict cloud thickness and generating the prediction result of cloud thickness in the target area includes preprocessing and extracting the cloud data to be identified, inputting it into the trained artificial intelligence model, the model outputting the predicted value of cloud thickness, establishing a feedback mechanism, and comparing the identification result with the actual observation data.

[0043] Collect the output data after cloud thickness identification, record the cloud thickness category of each prediction output, convert the category into the corresponding value, and obtain the corresponding actual cloud thickness observation data. Organize the prediction data and actual observation data into a paired dataset, organize it into a table, and measure the data bias through classification accuracy and confusion matrix.

[0044] Based on the analysis results, the GNN cloud thickness prediction model was optimized by adjusting the model structure, adding training data, or retraining the model.

[0045] Another objective of this invention is to provide a cloud thickness recognition system based on the joint Cycle-GAN and GNN, which can solve the shortcomings of existing methods in handling heterogeneous data sources and cloud thickness prediction accuracy by fusing multi-source data and modeling spatial-temporal relationships.

[0046] To solve the above technical problems, the present invention provides the following technical solution: a cloud thickness recognition system based on Cycle-GAN and GNN, comprising: a data acquisition and preprocessing module, a feature extraction module, a Cycle-GAN data conversion module, a graph neural network analysis module, and a model optimization and feedback mechanism module;

[0047] The data acquisition and preprocessing module is responsible for collecting meteorological data of the target area, including satellite data, radar data, and lidar data.

[0048] The feature extraction module extracts cloud features, including texture features, spectral features, geometric features, and features related to atmospheric environmental parameters. The extracted features will be used as input data for feature transformation and cloud thickness testing.

[0049] The Cycle-GAN data transformation module transforms the features of different data sources using Cycle-GAN, eliminates the differences between different data formats, constructs several generators and discriminators, performs style transformation on satellite, radar, and lidar data, optimizes the generator through adversarial training and cycle consistency loss, and determines the quality and consistency of the transformed data.

[0050] The graph neural network analysis module uses the data points transformed by Cycle-GAN as nodes in a graph, models the spatial and temporal relationships between data using the graph structure, and learns a predictive model for cloud thickness by combining graph convolution operations and node features.

[0051] The model optimization and feedback mechanism module evaluates the cloud thickness prediction results, compares the prediction results with the actual observation data to calculate evaluation indicators, adjusts the model's hyperparameters based on the evaluation results, and optimizes the structure and training strategy of Cycle-GAN and GNN models.

[0052] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the cloud thickness recognition method based on Cycle-GAN and GNN as described above.

[0053] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a cloud thickness recognition method based on Cycle-GAN and GNN as described above.

[0054] The beneficial effects of this invention are as follows: By conducting collaborative observations with multiple devices and comprehensively utilizing multi-source data such as meteorological satellites, ground-based radar, and lidar, this invention enhances the comprehensiveness and spatiotemporal synchronization of cloud data; by employing a unified data processing method, it improves data quality and availability, ensuring accuracy and reliability; by extracting multi-dimensional features, it enhances the ability to characterize clouds and provides richer information for cloud thickness identification; by leveraging the joint model architecture of Cycle-GAN and GNN, it optimizes data transformation and graph structure analysis, significantly improving identification accuracy and reliability, and through a continuous optimization mechanism, it ensures good model performance in constantly changing environments. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the 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.

[0056] Figure 1The first embodiment of the present invention provides an overall flowchart of a cloud thickness recognition method based on the joint Cycle-GAN and GNN. Detailed Implementation

[0057] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0058] Example 1, referring to Figure 1 As one embodiment of the present invention, a cloud thickness recognition method based on a joint Cycle-GAN and GNN is provided, comprising:

[0059] Utilizing various meteorological detection devices, including meteorological satellites, ground-based radar, and lidar, continuous data collection is conducted on the target area according to a pre-set observation plan and parameter settings. Meteorological satellites acquire multi-band cloud image data at specific time intervals; ground-based radar scans the surrounding airspace in real time to obtain cloud echo intensity data; lidar emits high-frequency laser pulses to collect scattering information from cloud particles. This data encompasses electromagnetic radiation information, echo intensity, and scattering characteristics across different bands. Simultaneously, atmospheric environmental parameter data, such as temperature, humidity, and air pressure, are obtained from nearby meteorological stations, ensuring the temporal and spatial synchronization of all data to comprehensively reflect the spatiotemporal variation characteristics of clouds.

[0060] Electromagnetic radiation information in different bands encompasses multiple bands of data acquired by meteorological satellites and other equipment, including the visible light band and the infrared band. This electromagnetic radiation information can reflect the physical characteristics of clouds, such as the reflection, absorption, and emission of cloud layers. Different types of clouds exhibit different radiation characteristics in different bands, which can be used to identify information such as cloud type and altitude.

[0061] Echo intensity is primarily obtained through ground-based radar and lidar. When electromagnetic waves emitted by radar encounter cloud particles, they produce echoes. The echo intensity is related to factors such as the size, number, and distribution of particles within the cloud, and can help determine the structure and density of the cloud.

[0062] Scattering characteristics are collected by radar and lidar. The scattering characteristics of cloud particles to electromagnetic waves can provide microscopic physical information about cloud particles, such as their shape and phase (liquid, solid, etc.). Different scattering characteristics help to further analyze the properties of clouds.

[0063] Temperature: The vertical and horizontal distribution of atmospheric temperature has a significant impact on cloud formation, development, and dissipation. Temperature variations at different altitudes affect the condensation and evaporation processes of water vapor, thus influencing cloud thickness, height, and other characteristics. Surface temperature data can be obtained from meteorological stations, while upper-air temperature data can be obtained through radiosondes, meteorological satellite infrared remote sensing, and other methods.

[0064] Humidity: Water vapor is one of the key elements in cloud formation. The level of air humidity determines the water vapor content, and humidity data is crucial for understanding the water vapor saturation within clouds and their precipitation potential. Humidity data can be obtained through humidity sensors (such as hygrometers at ground weather stations) or by satellite remote sensing.

[0065] The distribution of air pressure is closely related to atmospheric motion and vertical structure. Changes in air pressure affect the rising and sinking of air, which in turn affects cloud formation and evolution. Air pressure data are typically measured by ground-based meteorological stations and upper-air meteorological sounding equipment (such as radiosondes).

[0066] The collected multi-source heterogeneous data were formatted and calibrated. Data standardization was performed based on the characteristics of data collected from different devices, mapping the data to a unified numerical range and coordinate system.

[0067] Identify the data formats from different data sources. Organize the data formats collected from various devices, including meteorological satellites, ground-based radar, lidar, and various weather stations. Meteorological satellite data is stored in specific satellite data formats (HDF, NetCDF), containing different levels of file structure and data organization, recording multi-band radiation information of clouds, etc. Ground-based radar data may be saved in binary file format, with its data structure organized around parameters such as radar echo intensity. Atmospheric environmental parameters (temperature, humidity, air pressure) recorded by weather stations are presented in text file format (CSV), with each line corresponding to the observation value at a different time.

[0068] Choose a common data format conversion method. Convert to NetCDF format uniformly. Utilize the netCDF4 library in Python to process NetCDF format files and perform the format conversion operation.

[0069] Numerical range calibration. Min-max normalization is used to map the data to a uniform numerical range. The formula is:

[0070]

[0071] Where x is the original data, x max and x min These are the minimum and maximum values ​​of the original data, new max and new min The minimum and maximum values ​​within the target numerical range, xnew It is the new data value after normalization.

[0072] Coordinate system calibration. Ground-based radar data needs to be converted from the spatial location of clouds detected by the radar into latitude and longitude coordinates based on the radar's geographical location (latitude and longitude), detection direction, and scanning angle, using spatial geometric transformation formulas. The pyproj library in Python is used to achieve accurate coordinate transformation, ensuring that data acquired by different devices are consistent and matched in spatial location description.

[0073] Data cleaning removes outliers and noise caused by equipment malfunctions, electromagnetic interference, and other factors. Statistical analysis is used to determine the reasonable range of the data, and data points outside this range are eliminated. Missing values ​​are predicted using meteorological physical models to impute the missing data, ensuring data integrity.

[0074] Statistical analysis is used to determine reasonable ranges for each parameter and identify outliers. Observations from the same region at different times within a specific time period (e.g., the corresponding season over the past month or year) are collected. Statistical measures such as the mean, standard deviation, minimum, and maximum are calculated. Based on common data distribution patterns such as normality, a reasonable range is determined. This reasonable range is set as the mean ± n times the standard deviation; data points exceeding this range are considered outliers.

[0075] To remove noise, for continuous time-series data (such as radar echo intensity variations over time), a moving average filter is used. An appropriate window size is set (the number of time steps corresponding to 5 minutes), and the average value of the data within the window is used to replace the data point at the center of the window, smoothing out high-frequency small fluctuations in noise. For image-based data (such as satellite cloud images), a spatial filtering method is employed. A pixel neighborhood (such as a 3×3 pixel area) is selected, and the median value of the pixels within that neighborhood is used to replace the center pixel value. This removes abnormal pixels such as salt-and-pepper noise from the image, making the image clearer and smoother, facilitating subsequent feature extraction and analysis.

[0076] Predicting missing values ​​using meteorological physics models. By establishing relevant physical models such as cloud microphysics models and atmospheric circulation models, and using known relevant parameters as input (such as temperature, humidity, air pressure of the surrounding area, and partial cloud observation data), the models simulate atmospheric processes to calculate missing cloud-related data or atmospheric environmental parameter values. Taking the prediction of missing values ​​for cloud water vapor content (humidity-related data) as an example, if information such as cloud top height, cloud top temperature, and surrounding atmospheric pressure is known, the reasonable water vapor content in the cloud at the corresponding height can be calculated based on the water vapor saturation equation and the vertical thermal structure model within the cloud. This fills in the missing humidity data, ensuring data integrity and providing a complete data foundation for subsequent analyses such as cloud thickness identification.

[0077] Determine the relevant meteorological physical model and required input parameters. Based on the target parameter to be predicted (in this case, cloud water vapor content) and existing known data, select a suitable meteorological physical model. For cloud water vapor content prediction, water vapor saturation equations and vertical thermal structure models within the cloud are commonly used. Required input parameters include cloud top height (H). top Cloud top temperature (T) top The parameters include atmospheric pressure (P) and other relevant auxiliary parameters (such as air density, denoted as ρ). These parameters can be obtained from data collected by various devices such as meteorological satellites, ground-based radars, and weather stations, or through prior data processing and analysis.

[0078] Calculations related to the water vapor saturation equation. Calculating the saturated water vapor pressure (E) at the cloud top temperature using the water vapor saturation equation. s The Clausius-Clapeyrone equation is used, which gives the vapor pressure at a given temperature when water vapor reaches saturation. Its empirical formula (applicable to a certain temperature range, such as the temperature range commonly found in the troposphere) is as follows:

[0079]

[0080] In the formula, E s The unit is usually hectopascal (hPa), T top The unit is Celsius (°C).

[0081] Consider the vertical thermal structure and pressure variations within clouds. The vertical thermal structure within clouds affects the variation of water vapor content with altitude, while air pressure is also a key factor. The vertical pressure variations of the atmosphere are generally described using static equations, which are:

[0082]

[0083] Where z represents the height (as opposed to the cloud top height H) top (These are related and used for vertical integration calculations), where g is the acceleration due to gravity (approximately 9.8 m / s²). 2 By performing operations such as integration on the static equations, an expression for the change of air pressure with altitude can be obtained, and then combined with other conditions, the water vapor content at different altitudes can be calculated.

[0084] Assume the air pressure at the cloud top is known to be P. top (This can be obtained through actual measurement or calculation based on atmospheric environment data), from the cloud top downwards to the target height H (let's call it H). <H top The formula for calculating the pressure change at point (,) is as follows:

[0085]

[0086] In the formula, P(H) is the air pressure value at altitude H. In actual calculations, the air density ρ will change with factors such as altitude and temperature. Usually, its expression with altitude can be determined according to the atmospheric state equation (such as the ideal gas state equation P=ρRT, where R is the gas constant and T is the temperature), and then integral calculation can be performed.

[0087] The water vapor content can be estimated based on the relative humidity relationship.

[0088] After obtaining the air pressure and saturated water vapor pressure at the target altitude, the actual water vapor content can be calculated by combining the concept of relative humidity (RH) (which can be expressed in the form of water vapor density, denoted as ρ). v Relative humidity is defined as the difference between actual water vapor pressure (E) and saturated water vapor pressure (E). s The ratio of ), that is:

[0089]

[0090] And water vapor density (ρ v There is a certain physical relationship between the vapor pressure (R) and the water vapor pressure (E). Under the assumption of ideal gas conditions, this relationship can be established using the following formula (R). v (where the gas constant is for water vapor):

[0091]

[0092] Therefore, based on the known relative humidity (if relevant statistical data or empirical estimates are available) and the previously calculated saturated vapor pressure, the actual vapor pressure E = RH × E is first calculated. s Then, substituting these values ​​into the water vapor density calculation formula, we obtain the water vapor density within the cloud at the target altitude (which is a quantitative representation of water vapor content):

[0093]

[0094] Temperature T can be calculated at the target altitude using a thermal stratification model based on atmospheric stratification, based on a vertical thermal structure model within the cloud and information such as cloud top temperature and altitude changes.

[0095] Let the height of the tropopause be H. tropopause The temperature lapse rate within the troposphere is γ tropo The coefficient of temperature variation with altitude within the stratosphere is α. strato (Assuming the slope of the linear relationship between temperature and altitude within the stratosphere is positive, the temperature at the bottom of the stratosphere is T.) strato_base (i.e., the temperature corresponding to the top of the troposphere).

[0096] If the target altitude H is within the troposphere (H≤H) tropopause If the temperature calculation formula follows the same linear decreasing relationship as described above:

[0097] T = T top +γ tropo ×(HH top )

[0098] If the target altitude H is within the stratosphere (H>H) tropopause If the temperature of the stratosphere is such that the temperature of the stratosphere is calculated using the following formula:

[0099]

[0100] Fill in the missing humidity data. The water vapor content (in the form of water vapor density, etc.) data at the target altitude within the cloud, calculated through the above steps, is used to fill in the missing humidity data in the original positions. This completes the missing value filling process, ensuring the integrity of the humidity data in the dataset, so that it can be better applied to subsequent analysis steps such as cloud thickness identification.

[0101] Rich features are extracted from the preprocessed cloud data. These include cloud texture features, spectral features (ratios and differences of radiant intensities in different bands, etc.), geometric features (cloud shape, area, height, etc.), and features related to atmospheric environmental parameters (such as the difference between cloud top temperature and ambient temperature, and the impact of humidity on cloud boundaries). These features comprehensively characterize the properties of clouds, providing ample information for subsequent model training.

[0102] Texture features of clouds are extracted using gray-level co-occurrence matrix.

[0103] Determine several key parameters for calculating the gray-level co-occurrence matrix, including the displacement vector. It represents the spatial distance and orientation between pixel pairs. For example, d x =1,d y =0 indicates a pixel pair with a horizontal spacing of 1 pixel; the angle can typically be 0° (horizontal), 45°, or 90°.

[0104] (Vertical direction) and 135°. The number of gray levels n also needs to be determined. Generally, the gray range of the image can be quantized into n levels, such as n=8 or n=16, which helps reduce computation and highlight texture features.

[0105] The image is quantized. The original image's grayscale value range is [0, 255]. To quantize it into 8 grayscale levels, the grayscale range can be divided into 8 intervals, such as:

[0106] [0,31]→0, [32,63]→1, ..., [224,255]→7

[0107] In this way, the grayscale value of each pixel is remapped to a smaller integer range.

[0108] Calculate the elements of the co-occurrence matrix. Let the image be I(x,y), for a given displacement vector... And the number of gray levels n, gray-level co-occurrence matrix The element calculation method (where i and j are gray levels, ranging from 0 to n-1) is as follows: traverse each pixel (x, y) in the image, and check the pixel (x+d) x ,y+d y Is the pixel (x, y) within the image range? If it is within the range, and the gray level of pixel (x, y) is i, then the gray level of pixel (x+d) is i. x ,y+d y If the gray level of a given element is j, then the elements in the co-occurrence matrix will be... Add 1. After completing the traversal, the co-occurrence matrix will be... Each element in the matrix is ​​divided by the sum of all elements and normalized. The resulting co-occurrence matrix elements represent the probability of a pair of pixels with gray levels i and j occurring under a given displacement vector.

[0109] Further explanation regarding the extraction of texture features is as follows:

[0110] Contrast ratio measures the sharpness of textures and the abruptness of grayscale changes in an image. The calculation formula is:

[0111]

[0112] A higher contrast value indicates more dramatic changes in texture within the image.

[0113] Correlation is used to describe the linear dependence of pixel pairs in an image. The calculation formula is:

[0114]

[0115] Where μ i and μ j These are the mean values ​​in the row and column directions, respectively, σ i and σ j These represent the standard deviations in the row and column directions, respectively. A correlation value close to 1 or -1 indicates a strong linear relationship between pixel pairs, while a value close to 0 indicates a weaker linear relationship. In cloud maps, correlation can help determine the regularity of cloud textures; for example, stratiform clouds may have a high correlation.

[0116] Energy, also known as the angular second-moment, reflects the uniformity of image texture. The calculation formula is:

[0117]

[0118] The higher the energy value, the more uniform the texture. For example, large, uniform cloud layers have higher energy values, while complex and varied cloud textures have lower energy values.

[0119] Entropy is a measure of the complexity or randomness of image texture. The formula for calculation is:

[0120]

[0121] A higher entropy value indicates a more complex and irregular texture in the image. For example, broken cumulus cloud textures have a higher entropy value, while smooth cirrus cloud textures have a lower entropy value.

[0122] Further explanation regarding spectral feature extraction is as follows:

[0123] Calculation of the ratio of radiant intensity across different spectral bands. The ratio of radiant intensity across different spectral bands can highlight the relative differences in the reflection, absorption, or emission characteristics of clouds in different spectral bands. This relative difference helps to distinguish between different types of clouds, different cloud phases (such as water clouds and ice clouds), or different states of clouds (such as thick clouds and thin clouds), because different cloud materials and structures have different optical properties in different spectral bands.

[0124] The radiative intensity data of clouds in band b1 and band b2 were obtained from meteorological satellites or other spectroscopic measurement equipment, and are denoted as follows: and The formula for calculating the ratio of the radiation intensities of these two bands is:

[0125]

[0126] For example, for satellite remote sensing data, the ratio of radiant intensity in common visible light bands (such as the red band, with a center wavelength of approximately 660 nm) to near-infrared bands (such as 860 nm) is useful for identifying cloud vegetation cover (if there is vegetation beneath the cloud) or cloud type. If I 660nm It is the radiation intensity in the red light band, I 860nm If we consider the near-infrared radiation intensity, then their ratio... It can be used as a spectral feature. In practical applications, this ratio may vary due to factors such as cloud thickness, particle size and distribution within the cloud.

[0127] Calculation of radiation intensity differences across different spectral bands. Calculating the differences in radiation intensity across different spectral bands can more intuitively reflect the absolute differences in the absorption or reflection of clouds at different spectral bands. This difference is helpful in identifying specific substances within clouds (differences in absorption of ice crystals or liquid water at different spectral bands) or distinguishing the physical states of different clouds (thick clouds absorb radiation more strongly, leading to changes in the difference between different spectral bands).

[0128] Suppose we obtain the radiation intensity data of clouds in bands b3 and b4, denoted as... and The formula for calculating the difference in radiation intensity between these two bands is:

[0129]

[0130] For example, in the shortwave infrared band, ice clouds exhibit strong absorption characteristics at 1.38 μm, while absorption is relatively weaker at 1.2 μm. If I 1.38μm It is the radiation intensity in the 1.38μm band, I 1.2μm If the radiation intensity is in the 1.2μm band, then their difference is Difference = I 1.2μm -I 1.38μm This can be used as a spectral feature to help identify ice clouds. Furthermore, as cloud thickness increases, this difference may change due to multiple absorptions and scatterings of radiation by the cloud, thus providing clues for cloud thickness identification.

[0131] Calculation of the combined characteristics of band ratios and differences. Calculation of (Ratio1-Ratio2)

[0132] in,

[0133] Calculate (Difference1 × Ratio3)

[0134] in,

[0135] Combined features can more comprehensively reflect the spectral characteristics of clouds because they take into account the interactions and changes of clouds in multiple bands.

[0136] To ensure comparability of features calculated from different band combinations, normalized band ratios and differences can be performed. For band radiation intensity ratios, the normalized ratio is calculated as follows:

[0137]

[0138] For the difference in radiation intensity across bands, calculate the normalized difference:

[0139]

[0140] Normalized features can be better applied to subsequent data analysis and model training as input features, avoiding the impact on model performance due to large differences in data volume.

[0141] Geometric feature extraction:

[0142] Geometric features (cloud shape, area, height, etc.)

[0143] Cloud shape extraction. The Canny edge detection algorithm is used to obtain the cloud contours, detecting edges based on image grayscale changes, where cloud edges typically represent areas of significant grayscale variation. Gaussian filtering is applied to the cloud image to reduce noise impact on edge detection. Let the cloud image be I(x,y). The Gaussian-filtered image G(x,y) can be obtained by convolving it with the Gaussian kernel function Gaussian(x,y,σ), as shown in the formula:

[0144] G(x,y)=I(x,y)*Gaussian(x,y,σ)

[0145] Where σ is the standard deviation of the Gaussian kernel, and the convolution operation is performed across the entire image. The gradient magnitudes G of the filtered image in the horizontal and vertical directions are calculated. m And the gradient direction θ. The horizontal gradient G. x and vertical gradient G y It can be obtained by convolving G(x,y) with the Sobel operator (horizontal and vertical directions) respectively.

[0146] Gradient magnitude:

[0147]

[0148] gradient direction Edge pixels are determined using non-maximum suppression and double thresholding to obtain the cloud outline. Non-maximum suppression checks whether a pixel is a local maximum along the gradient direction, while double thresholding determines strong and weak edges based on high and low thresholds. The weak edges are then connected to the strong edges to obtain the final outline.

[0149] Cloud area feature extraction. A pixel counting method is used. The cloud image is binarized, and the number of cloud pixels in the image is counted. This count can approximate the cloud area. Let the binary cloud image be B(x,y), and the cloud area A = ∑ x ∑ y B(x,y), where the summation is performed over the entire image. If the image resolution is known, for example, each pixel represents r × r square meters of actual ground, then the actual area of ​​the cloud is A × r. 2 square meters.

[0150] Cloud height features are extracted using a combination of satellite remote sensing data and meteorological models. Satellite remote sensing is used to obtain cloud top temperature information, which is then combined with an atmospheric temperature vertical profile model (such as a standard atmospheric model or a local atmospheric model obtained from sounding data) to estimate cloud height. Because atmospheric temperature typically varies with altitude, the difference between cloud top temperature and the surrounding atmospheric temperature can help determine cloud top height. First, the cloud top temperature T is obtained from satellite data. top By using the atmospheric temperature vertical profile model T(h) (where h is altitude), we find a model that satisfies T(h) = T top The height value h is the cloud top height.

[0151] For example, in a simple linear temperature decrease model:

[0152] T(h)=T0-γh

[0153] (T0 is the ground temperature, γ is the temperature lapse rate) in the following:

[0154] Cloudtop height:

[0155]

[0156] To estimate cloud base height, we can combine cloud type, humidity and other information, and estimate it by assuming the temperature and humidity distribution inside the cloud. For example, we can assume that the temperature inside the cloud is approximately the cloud top temperature, and estimate the cloud base height based on the water vapor saturation temperature-altitude relationship.

[0157] Characteristics related to atmospheric environmental parameters:

[0158] Extraction of the difference between cloud top temperature and ambient temperature.

[0159] Obtain the cloud top temperature T ct and the ambient temperature T around the cloud env Then, the difference between the cloud top temperature and the ambient temperature:

[0160] ΔT=T ct -T env

[0161] This difference can be used as a feature to describe the difference in thermal state between the cloud and the surrounding atmosphere. During the development of convective clouds, a large difference between the cloud top temperature and the ambient temperature indicates strong convective activity within the cloud and vigorous cloud development.

[0162] Feature extraction of the effect of humidity on cloud boundaries.

[0163] Let one side of the cloud boundary be the inside of the cloud (humidity H). in On the other side is outside the clouds (humidity is H). out Humidity gradient:

[0164]

[0165] Here, Δd represents the distance interval between the two sides of the cloud boundary. The humidity gradient reflects the direction and intensity of water vapor transport at the cloud boundary; a larger humidity gradient may indicate that the cloud is growing through water vapor condensation or dissipating due to evaporation. Correlation analysis between humidity and cloud boundary shape is performed to statistically analyze the shape changes of the cloud boundary under different humidity conditions. For example, when humidity is high, the cloud boundary may be more blurred because the diffusion of water vapor widens the transition area between the cloud and the surrounding atmosphere; while at lower humidity, the cloud boundary may be relatively clear. By analyzing a large number of cloud images and corresponding humidity data, a statistical relationship can be established between humidity and cloud boundary shape characteristics (such as boundary curvature and roughness). For example, the correlation coefficient r between the cloud boundary curvature C and humidity H can be calculated. If r has a significant value (close to 1 or -1), it indicates a strong correlation between humidity and cloud boundary curvature.

[0166] Cloud thickness recognition is achieved by combining Cycle-GAN and GNN, and a sample set is formed based on the collected data and extracted features.

[0167] Electromagnetic radiation data in different bands obtained from meteorological satellites are denoted as... Where i and j represent the spatial dimensions of the data (e.g., rows and columns in an image), each element It is a vector containing radiation values ​​of multiple bands.

[0168] Echo intensity data obtained from ground-based radar, denoted as... Similarly, i and j represent spatial dimensions.

[0169] The scattering characteristic data obtained from the lidar is denoted as...

[0170] Atmospheric environmental parameter data are obtained from meteorological stations, including temperature T = {t k Humidity H={h k} and air pressure P={p k}, where k represents different weather station locations or time points.

[0171] Extracting feature data, texture features include contrast, correlation, energy, and entropy; spectral features include normalized ratio and normalized difference; geometric features include cloud shape G(x,y), cloud area B(x,y), and cloud height T(h); atmospheric environmental parameters include the difference between cloud top temperature and ambient temperature ΔT; and the influence of humidity on cloud boundaries. Synchronously included in the training set, denoted as:

[0172]

[0173] Normalize and align the data.

[0174] For satellite electromagnetic radiation data, normalization is performed separately for each band. Let the data for a certain band be s. band The normalization formula is:

[0175]

[0176] Perform this normalization process on all band data to obtain normalized satellite data. The radar echo intensity data and lidar scattering characteristic data were also normalized to obtain... and For atmospheric environmental parameters, temperature, humidity, and air pressure are normalized separately. For example, temperature normalization: The normalized temperature T is obtained norm Humidity H norm and air pressure P norm .

[0177] Data from different data sources may not perfectly match spatially and temporally, aligning them is necessary. When satellite and radar data have different spatial resolutions, interpolation or other methods can be used to adjust them to the same resolution. Finally, data from different data sources can be integrated into a single comprehensive dataset.

[0178]

[0179] Cycle-GAN model construction and training, and the construction of Cycle-GAN generator.

[0180] Construct three generators, G 12 Converting satellite data and atmospheric environmental parameters into a form similar to radar data, G 23 Converting radar data and atmospheric environmental parameters into a format similar to lidar data, G 31 The LiDAR data and atmospheric environmental parameters are converted into a form similar to satellite data. The generator employs a multi-layer convolutional neural network (CNN) structure, containing multiple convolutional layers, batch normalization layers, and activation functions; this patent uses ReLU.

[0181] Discriminators: Construct three discriminators D1, D2, and D3. 3, These are used to distinguish between genuine and fake satellite data, radar data, and lidar data, respectively. The discriminator adopts a CNN structure.

[0182] The generator's total loss is a weighted sum of the adversarial loss and the cycle consistency loss. Let G...12 For example, the total loss is calculated using the following formula:

[0183]

[0184] Where, λ cyc These are the weights of the cycle consistency loss. The discriminator's loss is the adversarial loss.

[0185] The training process uses the Adam optimizer to alternately train the generator and discriminator. First, the generator is fixed while the discriminator is updated, then the discriminator is fixed while the generator is updated. This process is repeated for a certain number of rounds until the model converges.

[0186] Construction and training of graph neural networks (GNNs).

[0187] The graph is constructed by defining nodes. Data points transformed by Cycle-GAN (including satellite, radar, lidar data formats, and atmospheric environmental parameters) are used as nodes. Each node's feature vector contains its corresponding data value and spatial location information. For example, the feature vector of a node n can be represented as:

[0188] f n =[g 12 (D norm ) ij ,g 23 (D norm ) ij ,g 31 (D norm ) ij ]

[0189] Among them, g 12 g 23 g 31 The data is transformed by the Cycle-GAN generator, where i and j are spatial coordinates.

[0190] Edges are established based on spatial distance and data relevance. An edge is established between two nodes if the spatial distance between them is less than a certain threshold, or if the similarity of their feature vectors (such as radiation values ​​in different bands, echo intensity, etc.) is greater than a certain threshold (which can be calculated using methods such as cosine similarity).

[0191] It should be further explained that:

[0192] By using data points transformed by Cycle-GAN as nodes and establishing edges based on spatial distance and data correlation, the relationships between data points are clarified, providing a structured foundation for data processing using graph convolutional networks. The constructed graph can then be further processed using graph convolutional networks. Graph structures can effectively capture the spatial relationships and data correlations between data points, uncovering potential patterns and regularities in the data.

[0193] The output is a graph structure where each node contains a feature vector consisting of its corresponding data value and spatial location information, and edges are established between nodes based on spatial distance and data correlation. This graph structure serves as the input data format for a graph convolutional network.

[0194] Choosing a Graph Convolutional Network (GCN), the message passing formula is:

[0195]

[0196] Among them, H l It is the node feature matrix of the l-th layer. (A is the adjacency matrix, I is the identity matrix) yes The degree matrix, W l σ is the weight matrix of the l-th layer, and σ is the activation function, using ReLU.

[0197] Define the output layer and loss function. Add a linear layer to the last layer of the GNN to output the cloud thickness prediction. The loss function is the mean squared error (MSE).

[0198]

[0199] Among them, y i This is the actual cloud thickness value. This is the predicted cloud thickness value, and N is the number of samples.

[0200] It should be further explained that:

[0201] Cycle-GAN outputs the training data, and GNN uses the training data to output the final cloud thickness prediction result.

[0202] The GNN is trained using training data, and its parameters are tuned by minimizing the loss function using the Adam optimizer. During training, a validation set can be used for model selection and hyperparameter tuning.

[0203] An alternating training approach is employed. First, the Cycle-GAN is trained for several epochs. Then, the GNN is trained with the Cycle-GAN parameters fixed. Afterward, based on the GNN's performance on the validation set, the weights of the Cycle-GAN generator and discriminator are adjusted appropriately, and the GNN is trained again. This process is repeated until the joint model achieves satisfactory performance on the test set. The performance of the joint model is evaluated using the test set, and evaluation metrics (such as mean squared error, mean absolute error, etc.) are calculated. Based on the evaluation results, the model's hyperparameters are adjusted, such as the number of layers, kernel size, and learning rate of the Cycle-GAN generator and discriminator, as well as the number of layers, number of hidden units, and edge construction threshold of the GNN, to further improve the model's performance.

[0204] After preprocessing and feature extraction, the cloud data to be identified is input into a trained artificial intelligence model, which outputs a predicted value of cloud thickness.

[0205] For the regression model, the cloud thickness value is directly output. The trained regression model is f. reg (x) is the integrated feature vector x input into the model. The model directly outputs the predicted cloud thickness. Its calculation formula is the forward propagation calculation process of the model, which can generally be expressed as:

[0206] f reg (x)=w T x+b

[0207] Where w is the weight vector learned by the model, x is the input feature vector, and b is the bias term, then the predicted value... The result is calculated based on this linear relationship.

[0208] For the classification model, the thickness category to which the cloud belongs is output based on the pre-defined cloud thickness categories.

[0209] The classification model outputs a probability distribution of the thickness category to which the cloud belongs, using the softmax function to achieve category probability transformation. Let the class probability vector output by the model be p = [p1, p2, ..., p...]. C ], where C is the pre-defined number of cloud thickness categories, p k The probability that a cloud belongs to the k-th cloud type is expressed by the following formula:

[0210]

[0211] Among them, z k This is the unnormalized score corresponding to the k-th class, calculated by the classification model in the last layer (usually a value obtained after feature learning and transformation from previous layers, related to the input feature vector x and model parameters). The final predicted cloud thickness class is the class with the highest probability, i.e.:

[0212]

[0213] In other words, the category index k corresponding to the element with the largest value in the probability vector p is selected as the predicted cloud thickness category.

[0214] Establish a feedback mechanism to compare the identification results with actual observation data. Collect the output data after cloud thickness identification. Record the cloud thickness category of each prediction output, denoted as . At the same time, categories can be converted into corresponding numerical representations and uniformly denoted as... Obtain actual observation data. Obtain the actual observation data of cloud thickness corresponding from reliable sources such as high-altitude balloon measurements and professional meteorological station observations, and denote it as y i . Organize the prediction result data and the actual observation data into a paired data set. For example, it can be organized in tabular form, with each row containing the predicted value corresponding to one prediction and the actual observation value y i .

[0215] Consider specific metrics such as classification accuracy to measure the deviation. Accuracy: Calculate the proportion of samples with correct predictions in the total number of samples. The formula is:

[0216]

[0217] where I(·) is the indicator function. When the condition inside the parentheses holds (i.e., the predicted class is equal to the actual class c i ), the function value is 1; otherwise, it is 0.

[0218] Confusion Matrix: Construct a confusion matrix to analyze the misclassification situation between various classes in detail. The confusion matrix is a C×C matrix (C is the number of classes), and its element CM pq represents the number of samples that actually belong to class p but are predicted as class q. Through the confusion matrix, it can be intuitively seen which classes are prone to confusion, and then the deviation performance of the model on different classes can be analyzed. <l

[0219] If it is found that there is a deviation in the recognition result, analyze the reasons for the deviation, which may be errors in the data collection link, limitations of the model, or special changes in the atmospheric environment, etc.

[0220] Optimize the model according to the analysis results, adjust the model structure, increase the training data, or retrain the model to continuously improve the accuracy and reliability of cloud thickness recognition. If the deviation metrics (such as MSE or classification accuracy) of the model perform well on the training set but have a large deviation on the test set or actual verification data (such as a significant increase in MSE or a large decrease in accuracy), the model complexity can be reduced by reducing the number of layers or neurons of the model. For example, for a multi-layer neural network, assume the original model has L layers, and the number of neurons in each layer is n l (l = 1, 2,..., L), and try to reduce the number of layers to L′ (L′ < L), or adjust the number of neurons in some layers to n′ l (n′ l < n l ). Calculate the change in the number of model parameters after adjustment (taking the fully connected layer as an example): The formula for calculating the number of parameters of the fully connected layer is It is the number of input neurons, n out This refers to the number of output neurons. If one layer is removed, the difference in the number of parameters before and after the layer reduction is used to measure the reduction in model complexity. Assume that the k-th layer is removed, and its input neurons are n. k-1 The output neuron is n k The parameter reduction is n. k-1 ×n k .

[0221] Simultaneously using the cross-entropy loss function, for C categories, let the predicted probability distribution be:

[0222] The actual category labels are y = [y1, y2, ..., y C (Usually one-hot encoding is used, that is, if it actually belongs to class k, then y) k =1, others are 0),

[0223] The loss function formula is: Similarly, you can use optimizers like Adam to update the parameters using similar steps until they are normal.

[0224] Example 2, an embodiment of the present invention, provides a cloud thickness recognition system based on Cycle-GAN and GNN joint, including: a data acquisition and preprocessing module, a feature extraction module, a Cycle-GAN data conversion module, a graph neural network analysis module, and a model optimization and feedback mechanism module.

[0225] The data acquisition and preprocessing module is responsible for collecting meteorological data of the target area, including satellite data, radar data, and lidar data.

[0226] The feature extraction module extracts cloud features, including texture features, spectral features, geometric features, and features related to atmospheric environmental parameters. The extracted features will be used as input data for feature transformation and cloud thickness testing.

[0227] The Cycle-GAN data transformation module transforms the features of different data sources using Cycle-GAN, eliminates the differences between different data formats, constructs several generators and discriminators, performs style transformation on satellite, radar, and lidar data, optimizes the generator through adversarial training and cycle consistency loss, and determines the quality and consistency of the transformed data.

[0228] The graph neural network analysis module uses the data points transformed by Cycle-GAN as nodes in a graph, models the spatial and temporal relationships between data using the graph structure, and learns a predictive model for cloud thickness by combining graph convolution operations and node features.

[0229] The model optimization and feedback mechanism module evaluates the cloud thickness prediction results, compares the prediction results with the actual observation data to calculate evaluation indicators, adjusts the model's hyperparameters based on the evaluation results, and optimizes the structure and training strategy of Cycle-GAN and GNN models.

[0230] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0231] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0232] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0233] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0234] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A cloud thickness recognition method based on a combination of Cycle-GAN and GNN, characterized in that, include: Use sensors to collect meteorological data of the target area and perform preprocessing; Feature extraction based on preprocessed data; Cycle-GAN is used to transform the extracted features, eliminating format differences between data sources; Graph neural networks are used to analyze the transformed data and extract the spatial and temporal relationships of cloud data. The GNN model is used to predict cloud thickness and generate prediction results for the cloud thickness of the target area.

2. The cloud thickness recognition method based on the joint Cycle-GAN and GNN as described in claim 1, characterized in that: The use of sensors to collect meteorological data of the target area includes: Using meteorological satellites, ground-based radar, and lidar, temperature data is continuously collected from the target area according to the preset observation plan and parameter settings. Meteorological data, including humidity and air pressure, are collected for the target area. The preprocessing involves unifying and calibrating the format of the collected multi-source heterogeneous data. Based on the characteristics of data collected from different devices, data standardization is performed, mapping the data to a unified numerical range and coordinate system, and identifying and converting data formats from different data sources into NetCDF format.

3. The cloud thickness recognition method based on the joint Cycle-GAN and GNN as described in claim 2, characterized in that: The feature extraction based on preprocessed data includes cloud texture features, spectral features, geometric features, and features related to atmospheric environmental parameters.

4. The cloud thickness recognition method based on the joint Cycle-GAN and GNN as described in claim 3, characterized in that: The texture features of the cloud are extracted using the gray-level co-occurrence matrix, and the key parameters for calculating the gray-level co-occurrence matrix are determined. The displacement vector is represented as follows: Where, d x d y These are the spatial distance and orientation between pixel pairs, respectively; By determining the number of gray levels n, the gray range of the image is quantized into n levels, reducing the amount of computation and highlighting texture features. The image is quantized, and the gray value of each pixel is remapped to a small integer range according to the number of gray levels n. The spectral features are calculated by taking the ratio of radiant intensity in each band and the difference in radiant intensity in each band, and then performing a combined feature calculation. The geometric features are extracted by creating the shape of the clouds, obtaining the cloud contours using the Canny operator edge detection algorithm, detecting edges based on image grayscale changes, applying Gaussian filtering to the cloud image, and calculating the gradient magnitude G in the horizontal and vertical directions. m The edge pixels are determined by non-maximum suppression and double thresholding based on the gradient direction θ, thus obtaining the outline of the cloud. The area features of clouds are extracted, the cloud images are binarized using the pixel counting method, the number of cloud pixels in the image is counted, the temperature information of the cloud top is obtained using satellite remote sensing, and the height of the cloud is estimated by combining the atmospheric temperature vertical profile model. The features related to the atmospheric environmental parameters are the difference between cloud top temperature and ambient temperature, and the influence of humidity on cloud boundaries.

5. The cloud thickness recognition method based on the joint Cycle-GAN and GNN as described in claim 4, characterized in that: The process of using Cycle-GAN to transform the extracted features involves forming a feature sample set based on the collected data and the extracted features, and then performing normalization processing. If the data do not match perfectly in space and time, alignment operations are performed. When satellite data and radar data have different spatial resolutions, they are adjusted to the same resolution using interpolation methods, integrating data from different data sources into a comprehensive dataset.

6. The cloud thickness recognition method based on the joint Cycle-GAN and GNN as described in claim 5, characterized in that: The transformation of extracted features using Cycle-GAN also includes constructing a Cycle-GAN generator, which consists of three generators G. 12 Converting satellite data and atmospheric environmental parameters into a form similar to radar data, G 23 Converting radar data and atmospheric environmental parameters into a format similar to lidar data, G 31 The generator uses a multi-layer convolutional neural network structure to convert lidar data and atmospheric environmental parameters into a form similar to satellite data. Construct three discriminators D1, D2, and D3 3, These are used to distinguish between genuine and fake satellite data, radar data, and lidar data, respectively. The generator's total loss is a weighted sum of the adversarial loss and the cycle consistency loss, expressed as: Among them, L G12 The generator's total loss is a weighted sum of the adversarial loss and the cycle consistency loss. Combating losses Cyclic consistency loss, D norm Integrate data from different data sources into a single comprehensive dataset. For normalized data (D) norm The probability distribution function of p, where r represents the real data sample. data (r) represents the probability distribution function of the real data (r), and discriminators D1, D2, and D3 are used to distinguish between real and fake satellite data, radar data, and lidar data, respectively. λ cyc It is the weight of the cycle consistency loss, and the loss of the discriminator is the adversarial loss; The training process uses the Adam optimizer to train the generator and discriminator alternately. First, the generator is fixed and the discriminator is updated. Then, the discriminator is fixed and the generator is updated. The training is iterated for a certain number of rounds until the model converges. The data points transformed by Cycle-GAN are used as nodes, and edges are established based on spatial distance and data correlation. A graph convolutional network is selected. Each node's feature vector contains its corresponding data value and its spatial location information; An edge is established if the spatial distance between two nodes is less than a threshold, or the similarity of their feature vectors is greater than a threshold. Define the output layer and loss function, and add a linear layer to the last layer of the GNN to output the cloud thickness prediction value; The loss function uses mean squared error, and its expression is: Among them, y i This is the actual cloud thickness value. This is the predicted cloud thickness value, where N is the number of samples; An alternating training approach is adopted. First, Cycle-GAN is trained for several rounds. Then, the parameters of Cycle-GAN are fixed to train GNN. Based on the performance of GNN on the validation set, the weights of the generator and discriminator of Cycle-GAN are adjusted. Then, GNN is trained again and this process is repeated until the joint model meets the requirements on the test set. The performance of the joint model is evaluated using a test set, and evaluation metrics are calculated. Based on the results, the hyperparameters of the GNN cloud thickness prediction model are adjusted to improve model performance.

7. The cloud thickness recognition method based on Cycle-GAN and GNN as described in claim 6, characterized in that: The method of using a GNN model to predict cloud thickness and generate prediction results of cloud thickness in the target area includes preprocessing and extracting features from the cloud data to be identified, inputting it into a trained artificial intelligence model, outputting the predicted cloud thickness value from the model, establishing a feedback mechanism, and comparing the identification results with the actual observation data. Collect the output data after cloud thickness identification, record the cloud thickness category of each prediction output, convert the category into the corresponding value, and obtain the corresponding actual cloud thickness observation data. Organize the prediction data and actual observation data into a paired dataset, organize it into a table, and measure the data bias through classification accuracy and confusion matrix. Based on the analysis results, the GNN cloud thickness prediction model was optimized by adjusting the model structure, adding training data, or retraining the model.

8. A system employing a cloud thickness recognition method based on a combined Cycle-GAN and GNN as described in any one of claims 1 to 7, characterized in that: It includes a data acquisition and preprocessing module, a feature extraction module, a Cycle-GAN data conversion module, a graph neural network analysis module, and a model optimization and feedback mechanism module; The data acquisition and preprocessing module is responsible for collecting meteorological data of the target area, including satellite data, radar data, and lidar data. The feature extraction module extracts cloud features, including texture features, spectral features, geometric features, and features related to atmospheric environmental parameters. The extracted features will be used as input data for feature transformation and cloud thickness testing. The Cycle-GAN data transformation module transforms the features of different data sources using Cycle-GAN, eliminates the differences between different data formats, constructs several generators and discriminators, performs style transformation on satellite, radar, and lidar data, optimizes the generator through adversarial training and cycle consistency loss, and determines the quality and consistency of the transformed data. The graph neural network analysis module uses the data points transformed by Cycle-GAN as nodes in a graph, models the spatial and temporal relationships between data using the graph structure, and learns a predictive model for cloud thickness by combining graph convolution operations and node features. The model optimization and feedback mechanism module evaluates the cloud thickness prediction results, compares the prediction results with the actual observation data to calculate evaluation indicators, adjusts the model's hyperparameters based on the evaluation results, and optimizes the structure and training strategy of Cycle-GAN and GNN models.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of any one of claims 1 to 7 for a cloud thickness recognition method based on Cycle-GAN and GNN.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of any one of claims 1 to 7 of the cloud thickness recognition method based on Cycle-GAN and GNN.

Images