Ground surface downlink short wave radiation space downscaling method based on super-resolution reconstruction technology

Abstract

The invention discloses an earth surface downlink short-wave radiation space downscaling method based on a super-resolution reconstruction technology, belongs to the technical field of meteorological data analysis, and solves the problems of insufficient solar radiation data space resolution and weak earth surface heterogeneity characterization capability in the prior art. Carrying out super-resolution reconstruction on the radiation data set by adopting a super-resolution combination model, and carrying out weighted fusion on a reconstructed data set output by the super-resolution combination model; according to the method, the output results of the multiple models are subjected to weighted fusion, weight optimization of the super-resolution sub-models is performed in combination with the high-resolution terrain factors, spatial downscaling from 5km to 1km resolution is finally realized, the spatial precision of the solar radiation data is remarkably improved on the basis of keeping the hour-level time resolution of the original data, and the accuracy of the solar radiation data is improved. And more earth surface heterogeneity information can be captured, and meanwhile, the high-time-resolution application requirement is met.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of meteorological data analysis, and particularly relates to a surface downward shortwave radiation spatial downscaling method based on super-resolution reconstruction technology. BACKGROUND

[0002] Surface downward shortwave radiation (DSR) is a key component of the Earth's surface energy balance, directly affecting climate change research, solar resource assessment, precision agriculture management, and ecological environment monitoring. High-precision solar radiation data is of great scientific and practical value for understanding the surface energy budget, optimizing renewable energy utilization, and guiding agricultural production.

[0003] However, the current mainstream surface solar radiation data sets, such as CERES, ERA5, etc., have a spatial resolution of 25-100 km, while the highest precision data set in existing research has a spatial resolution of only 5 km. This situation seriously restricts the fine development of related field research and application, mainly existing the following problems:

[0004] Insufficient spatial resolution: Existing data sets cannot accurately capture the local impact of small-scale environmental characteristics such as terrain undulations, vegetation cover differences, and urban heat island effects on solar radiation. In complex terrain areas, such as mountainous regions, solar radiation is significantly affected by factors such as mountain blocking, slope and aspect, and low-resolution data cannot capture these details, resulting in inaccurate depiction of solar radiation spatial distribution.

[0005] Weak representation of surface heterogeneity: At a spatial scale of 5 km, details such as the radiation difference between valleys and mountain tops, and the radiation distribution gradient between farmland and cities are often treated as homogeneous. This treatment obscures the actual radiation differences between different underlying surface types, making it difficult to meet the needs of fine-scale research and application of surface radiation, and making research based on these data lack accuracy in reflecting the true situation to some extent.

[0006] Traditional downscaling methods have the following limitations:

[0007] Statistical downscaling method: This method relies on historical observation data to establish statistical relationships, but in data sparse areas, due to the lack of sufficient sample data, its effect is often poor, making it difficult to accurately downscale, resulting in a significant reduction in the reliability and accuracy of the downscaling results.

[0008] Dynamic downscaling methods, while theoretically capable of providing high spatial resolution data, are computationally extremely expensive, requiring significant computational resources and time to run complex physical models. This makes large-scale application of dynamic downscaling methods difficult in practice, limiting their use in wide-area and frequently updated scenarios.

[0009] Spatiotemporal fusion models have difficulties in processing hourly high temporal resolution data, and cannot well balance the high resolution requirements of both time and space dimensions. They are difficult to meet the application scenarios that have high requirements for both temporal continuity and spatial accuracy, such as short-term solar power generation forecasting and precision agricultural meteorological services.

[0010] Therefore, to address the issues of insufficient spatial resolution of solar radiation data and weak ability to characterize surface heterogeneity in existing technologies, we propose a spatial downscaling method for surface downwave radiation based on super-resolution reconstruction technology. Summary of the Invention

[0011] The purpose of this invention is to address the shortcomings of existing technologies by providing a spatial downscaling method for surface downwave radiation based on super-resolution reconstruction technology, which solves the problems of insufficient spatial resolution of solar radiation data and weak ability to characterize surface heterogeneity in existing technologies.

[0012] This invention is implemented as follows: a spatial downscaling method for surface downwave radiation based on super-resolution reconstruction technology, comprising:

[0013] Acquire low-resolution surface downwave radiation data, preprocess the surface downwave radiation data, and output a standardized radiation dataset;

[0014] Collect historical datasets and pre-train a super-resolution ensemble model using these datasets.

[0015] The radiation dataset is loaded in real time, and a super-resolution ensemble model is used to perform super-resolution reconstruction of the radiation dataset to obtain at least one set of high-resolution reconstructed datasets.

[0016] Obtain the reconstructed dataset, identify the terrain factor information in the reconstructed dataset, and combine the ground observation station data and terrain factor information to optimize the weights of each super-resolution sub-model in the super-resolution combined model by minimizing the loss function;

[0017] The reconstructed dataset output by the super-resolution combined model is weighted and fused based on the weights of the super-resolution sub-models to generate the final high-resolution solar radiation data.

[0018] Preferably, the method for preprocessing downwave radiation data from the ground surface includes:

[0019] Acquire surface downwave radiation data, use the three-standard-deviation principle combined with an adaptive threshold to detect outliers in the surface downwave radiation data, mark outliers in the surface downwave radiation data, and output surface downwave radiation data after removing outliers;

[0020] Load the surface downwave shortwave radiation data after removing outliers, analyze the time series characteristics of the surface downwave shortwave radiation data, identify and remove data points affected by cloud cover, and obtain the surface downwave shortwave radiation data after cloud cover removal.

[0021] Obtain surface downwave radiation data after cloud masking removal, standardize the surface downwave radiation data using the maximum-minimum method, and perform spatiotemporal scale matching processing on the surface downwave radiation data to output a standardized radiation dataset.

[0022] Preferably, the method for outlier detection of surface downwave radiation data using the three-standard-deviation principle combined with an adaptive threshold includes:

[0023] Load surface downwave radiation data, identify the climate sub-regions corresponding to the surface downwave radiation data, where the climate sub-regions include plateaus, plains, coasts, and lakes, and obtain at least one set of sub-region radiation datasets;

[0024] A sliding spatiotemporal window is used to perform spatiotemporal window segmentation on the sub-region radiation dataset to obtain at least one set of dynamic spatiotemporal window segments. The mean and standard deviation of the dynamic spatiotemporal window segments in the sub-region radiation dataset are calculated, and the initial anomaly threshold interval of the dynamic spatiotemporal window segments is set using the three-standard-deviation principle.

[0025] The initial anomaly threshold range is represented as follows:

[0026] α0=μ i ±3σ i

[0027] Where α0 represents the initial anomaly threshold interval, μ i ,σ i These are the mean and standard deviation of the dynamic spatiotemporal window segment, respectively.

[0028] The size of the sliding spacetime window is represented as follows:

[0029] W c =W base ×(1+q W ·C i +κ·A i )

[0030] Among them, W c W base These represent the size of the sliding spacetime window and the base window size, respectively.W κ represents the weighting coefficients of the sliding spatiotemporal window and the terrain factor weighting coefficient, respectively, and C i A i These are the slope component and the aspect effect coefficient, respectively.

[0031] Identify atmospheric transmittance and solar zenith angle within a dynamic spatiotemporal window segment, and adjust the initial anomaly threshold range based on atmospheric transmittance and solar zenith angle to obtain the adjusted adaptive threshold.

[0032] The adaptive threshold is expressed as:

[0033]

[0034] Among them, W α For physical adjustment of thresholds, T0,θ,T s These are the solar coefficient, solar zenith angle, and atmospheric transmittance, respectively.

[0035] Anomaly detection is performed on the surface downwave radiation data within the dynamic spatiotemporal window based on the adjusted adaptive threshold. Anomalies in the surface downwave radiation data are marked, and the surface downwave radiation data after removing anomalies is output.

[0036] Preferably, the method for identifying and removing data points affected by clouds includes:

[0037] Obtain surface downwave radiation data after removing outliers, and preset cloud pollution judgment thresholds based on spatiotemporal context detection algorithms;

[0038] Analyze the time series characteristics of surface downwave radiation data, and determine the actual clear-sky radiation value based on the time series characteristics of surface downwave radiation data;

[0039] The radiation characteristic ratio is determined by combining the actual value of clear-sky radiation and the theoretical value of clear-sky radiation, and it is then used to determine whether the radiation characteristic ratio corresponding to the downwave radiation data at the ground surface exceeds the preset cloud pollution judgment threshold.

[0040] The formula for calculating the radiation characteristic ratio is as follows:

[0041]

[0042] in, D represents the ratio of radiation characteristics. sj D l These are the actual value of clear-sky radiation and the theoretical value of clear-sky radiation, respectively, T. ait Γ represents atmospheric transmittance and aerosol optical thickness, respectively.

[0043] If the radiation characteristic ratio corresponding to the downwave radiation data of the surface exceeds the preset cloud pollution judgment threshold, it is determined that there is cloud pollution in the downwave radiation data of the surface, and the downwave radiation data of the surface is marked as cloud pollution.

[0044] If the radiation characteristic ratio corresponding to the surface downwave radiation data does not exceed the preset cloud pollution judgment threshold, it is determined that the surface downwave radiation data does not have cloud pollution, and the surface downwave radiation data without cloud pollution is retained.

[0045] Acquire surface downwave radiation data marked by cloud pollution and calculate the optical thickness of the surface downwave radiation data;

[0046] The formula for calculating the optical thickness of surface downwave radiation data is as follows:

[0047]

[0048] Among them, H ef Optical thickness for downwave radiation data from the Earth's surface;

[0049] Determine whether the optical thickness of the downwave radiation data on the ground surface exceeds a preset thickness threshold. If the thickness exceeds the preset thickness threshold, remove data points affected by clouds.

[0050] If the thickness does not exceed the preset thickness threshold, the spatiotemporal co-interpolation method is used to interpolate and reconstruct the surface downwave shortwave radiation data to obtain the reconstructed surface downwave shortwave radiation data.

[0051] The integrated and reconstructed surface downwave radiation data, and the surface downwave radiation data without cloud pollution, are surface downwave radiation data after cloud masking removal.

[0052] Preferably, the method for interpolating and reconstructing surface downwave radiation data using spatiotemporal co-interpolation includes:

[0053] Load surface downwave radiometric data and construct a spatiotemporal reconstruction model based on the surface downwave radiometric data to reconstruct the surface downwave radiometric data;

[0054] Perform topographic and shadow corrections on the surface downwave radiometric data in the spatiotemporal reconstruction model, and output the spatiotemporal reconstruction model of the surface downwave radiometric data after topographic and shadow corrections;

[0055] Solve for the data weighting coefficients of the surface downwave shortwave radiation data in the spatiotemporal reconstruction model, update the spatiotemporal reconstruction model with the weighting coefficients of the surface downwave shortwave radiation data, and obtain the surface downwave shortwave radiation data based on spatiotemporal co-interpolation.

[0056] The spatiotemporal reconstruction model is represented as follows:

[0057]

[0058] in, These represent the surface downwave radiometric data based on spatiotemporal co-interpolation, the surface downwave radiometric data after topographic and shading correction, and the data for sample point m at time t, respectively. B represents the number of sample points m, and w represents the data for sample point m. m Here, θ represents the data weighting coefficients for the surface downwave shortwave radiation data, DSR(x,y,t) represents the raw surface downwave shortwave radiation data of sample point m at time t, and θ represents the data weighting coefficients for the surface downwave shortwave radiation data. p The angle between the slope normal and the direction of the sun; M(h) is the number of sample points m at a given spatiotemporal distance h. They are the spacetime positions x m ,(x m Topographically and shadow-corrected surface downwave radiometric data at +h).

[0059] Preferably, the method for obtaining a super-resolution ensemble model through pre-training on a historical dataset includes:

[0060] Obtain the historical dataset, insert the preset square wave noise into the historical dataset to obtain the historical dataset with inserted noise, and divide the historical dataset into training set and test set;

[0061] Load the pre-built super-resolution ensemble model, freeze the parallel sub-modules, feature interaction layer and multimodal joint representation layer, and pre-train the denoising network using the training set to obtain the initial super-resolution ensemble model;

[0062] Freeze the denoising network in the initial super-resolution combined model, use the training set to train the parallel sub-modules in the initial super-resolution combined model independently, and use the LAMB optimizer to optimize the hyperparameters of the parallel sub-modules during training, and output the independently trained super-resolution combined model.

[0063] Initialize the feature interaction layer and preset the maximum likelihood node. Use the training set to jointly train the super-resolution combined model. Adjust the weights of the super-resolution sub-models during joint training to obtain the weight-adjusted super-resolution combined model.

[0064] Obtain a test set, use the test set as input, execute the super-resolution combined model, output the test results, compare the test results with the ground observation data, calculate the root mean square error, mean absolute error and correlation coefficient statistical indicators, and determine whether the root mean square error, mean absolute error and correlation coefficient statistical indicators all meet the preset index thresholds. If the root mean square error, mean absolute error and correlation coefficient statistical indicators all meet the preset index thresholds, output the converged super-resolution combined model.

[0065] Preferably, the super-resolution combined model is based on a parallel sub-module architecture. The parallel sub-module includes at least one set of super-resolution sub-models, such as ESRGAN, EDSR, VDSR, and RCAN models. The super-resolution combined model also includes an input layer, a denoising network, and an output layer. The input layer is connected to the denoising network, which is also connected to the parallel sub-module. The denoising network consists of an autoencoder and a decoder. A cross-model feature interaction layer and a multimodal joint representation layer are provided between the parallel sub-module and the output layer. The multimodal joint representation layer is a convolutional FConv layer. The denoising network is based on Barthelium... The Voss filter performs noise reduction on the radiation dataset and identifies noise autoencoders within it. An improved residual neural network removes residual noise interference from the radiation dataset, completing its reconstruction. The super-resolution sub-model is used for super-resolution reconstruction of the radiation dataset. Maximum likelihood estimation is introduced in the cross-model feature interaction layer. Based on maximum likelihood estimation, feature exchange nodes are defined during super-resolution reconstruction of the radiation dataset. The multimodal joint representation layer weights and fuses the reconstructed dataset output by the super-resolution combined model based on the weights of the super-resolution sub-model, generating the final high-resolution solar radiation data.

[0066] Preferably, when using a super-resolution combined model to perform super-resolution reconstruction of the radiation dataset, the super-resolution reconstruction process is represented as follows:

[0067]

[0068] Among them, f i Let i represent the i-th super-resolution sub-model. It is the 1km simulation result output by the i-th super-resolution sub-model;

[0069] The reconstructed dataset output by the super-resolution combined model is weighted and fused based on the weights of the super-resolution sub-models. The weighted fusion process is represented as follows:

[0070]

[0071] Among them, w i is the weight coefficient of the i-th super-resolution sub-model, and n is the total number of super-resolution sub-models.

[0072] Compared with the prior art, the embodiments of this application have the following main advantages:

[0073] In this embodiment of the invention, the output results of multiple models are weighted and fused, and the weights of the super-resolution sub-model are optimized by combining high-resolution terrain factors. This achieves spatial downscaling from 5km to 1km resolution. While maintaining the hourly temporal resolution of the original data, the spatial accuracy of solar radiation data is significantly improved. This allows for the capture of more surface heterogeneity information while meeting the needs of high temporal resolution applications, providing high-quality data support for climate research, solar energy resource assessment, precision agriculture, and other fields.

[0074] In this embodiment of the invention, during the preprocessing of surface downwave radiation data, a three-standard-deviation principle combined with an adaptive threshold is used to detect outliers in the surface downwave radiation data. This effectively identifies and removes outliers from the data. Furthermore, the dynamic threshold method can adapt to the characteristics of different climatic sub-regions, ensuring the accuracy and adaptability of outlier detection. The data after outlier removal is more accurate and reliable, reducing errors and biases caused by outliers. Identifying and removing data points affected by clouds effectively reduces the interference of clouds on radiation data, overcoming the problem of abnormally low radiation values ​​caused by the significant shading effect of clouds on solar radiation. By removing these cloud-affected data points, data that more accurately reflects the surface downwave radiation situation can be obtained, improving the accuracy and usability of the data.

[0075] In this embodiment of the invention, when using the three-standard-deviation principle combined with an adaptive threshold to detect outliers in surface downwave radiation data, a sliding spatiotemporal window is used to segment the sub-region radiation dataset into spatiotemporal windows. The initial outlier threshold interval is adjusted based on atmospheric transmittance and solar zenith angle to obtain an adjusted adaptive threshold. This significantly improves the accuracy and reliability of outlier detection in surface downwave radiation data. Furthermore, by integrating statistical laws and atmospheric physical mechanisms, it overcomes the limitations of traditional methods in areas with complex terrain and special weather conditions. The adaptive threshold adjustment mechanism allows for dynamic adjustment of the outlier threshold based on factors such as real-time atmospheric transmittance and solar zenith angle, without the need to pre-set a fixed threshold. This dynamic threshold adjustment capability enables the method to better cope with various complex environmental conditions and data changes, enhancing the method's adaptability and flexibility.

[0076] In this embodiment of the invention, when identifying and removing data points affected by clouds, the spatiotemporal context detection algorithm and the preset cloud pollution judgment threshold are combined to accurately identify data points affected by clouds. At the same time, the spatiotemporal collaborative interpolation method can comprehensively consider the correlation of data in time and space, which can not only retain more data points and avoid excessive data loss, but also ensure the continuity and consistency of the reconstructed data in time and space, thereby enhancing the integrity of the data. Attached Figure Description

[0077] Figure 1 This is a schematic diagram illustrating the implementation process of the surface downwave radiation spatial downscaling method based on super-resolution reconstruction technology provided by the present invention.

[0078] Figure 2 A schematic diagram of the implementation process of the preprocessing method for downwave radiation data from the Earth's surface is shown.

[0079] Figure 3 The diagram illustrates the implementation process of an outlier detection method for surface downwave radiation data using the three-standard-deviation principle combined with an adaptive threshold.

[0080] Figure 4 A schematic diagram illustrating the process of identifying and removing data points affected by clouds is shown.

[0081] Figure 5 The diagram illustrates the process of interpolating and reconstructing surface downwave radiation data using a spatiotemporal co-interpolation method.

[0082] Figure 6 The diagram illustrates the process of super-resolution reconstruction of a radiation dataset using a super-resolution combined model.

[0083] Figure 7 A schematic diagram illustrates the process of quality assessment and labeling of the generated high-resolution data. Detailed Implementation

[0084] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0085] To address the issues of insufficient spatial resolution and weak surface heterogeneity characterization in existing solar radiation data technologies, we propose a surface downscaling method for shortwave radiation based on super-resolution reconstruction. In short, the method first preprocesses the surface downscaling shortwave radiation data, then pre-trains a super-resolution combined model using historical datasets, and then uses this model to perform super-resolution reconstruction of the radiation dataset. The weights of each super-resolution sub-model in the super-resolution combined model are optimized by minimizing the loss function. Finally, the reconstructed dataset output by the super-resolution combined model is weighted and fused based on the weights of the super-resolution sub-models to generate the final high-resolution solar radiation data. In this embodiment, by weighted fusion of the outputs of multiple models and combining high-resolution terrain factors for weight optimization of the super-resolution sub-models, spatial downscaling from 5km to 1km resolution is achieved. While maintaining the hourly temporal resolution of the original data, the spatial accuracy of solar radiation data is significantly improved. This method can capture more surface heterogeneity information while meeting the needs of high temporal resolution applications, providing high-quality data support for climate research, solar energy resource assessment, precision agriculture, and other fields.

[0086] This invention provides a spatial downscaling method for surface downwave shortwave radiation based on super-resolution reconstruction technology. Figure 1 This diagram illustrates the implementation process of a spatial downscaling method for surface downwave radiation based on super-resolution reconstruction technology. The method specifically includes:

[0087] S10: Acquire low-resolution surface downwave radiation data, preprocess the surface downwave radiation data, and output a standardized radiation dataset.

[0088] In the preprocessing of surface downwave radiometric data, the low-resolution (5km) surface downwave radiometric data DSR is first read, and then data quality control is performed, including outlier detection and cloud masking. Finally, data standardization and format conversion are performed to ensure data consistency.

[0089] S20: Collect historical datasets and pre-train the super-resolution combined model using the historical datasets;

[0090] S30: Load the radiation dataset in real time, and use a super-resolution combined model to perform super-resolution reconstruction of the radiation dataset to obtain at least one set of high-resolution reconstructed datasets.

[0091] The super-resolution combined model includes at least one set of super-resolution sub-models, which include, but are not limited to, ESRGAN (Enhanced Super-Resolution Generative Adversarial Networks), EDSR (Enhanced Deep Super-Resolution), VDSR (Very Deep Super-Resolution), RCAN (Residual Channel Attention Networks), RDN (Residual Dense Network), and CARN (Cascading Residual Network). Figure 1 The coarse solar radiation product for the central region and the downscaled solar radiation product for the region are respectively low-resolution surface downwave radiation data and high-resolution solar radiation data.

[0092] S40, acquire the reconstructed dataset, identify the terrain factor information in the reconstructed dataset, combine the ground observation station data and terrain factor information, and optimize the weights of each super-resolution sub-model in the super-resolution combined model by minimizing the loss function; wherein, the terrain factors include digital elevation model (DEM), slope and aspect data. In this embodiment, the terrain factors are used as auxiliary information and constraints for weight optimization, and the resolution of the terrain factor data is high (≤1km).

[0093] In identifying and reconstructing topographic factors in the dataset, 1km resolution digital elevation model (DEM) data of the study area was acquired. Relief, location, slope, and aspect information were calculated using topographic analysis algorithms. The maximum rate of descent method was used for slope calculation, reflecting the degree of surface inclination. The gradient direction method was used for aspect calculation, representing the azimuth angle of the slope. Topographic factor preprocessing included: normalizing elevation values ​​according to the study area; converting slope values ​​to standard angle values ​​within the 0-90 degree range; and converting aspect values ​​to azimuth angles within the 0-360 degree range. These topographic factors will serve as important auxiliary information for subsequent weight optimization.

[0094] Then, the reference weights for each model are calculated based on the terrain complexity. First, the Terrain Complexity Index (TCI) is calculated. The TCI is categorized as extremely high (TCI ≥ 0.8), high (0.6 ≤ TCI < 0.8), medium (0.4 ≤ TCI < 0.6), low (0.2 ≤ TCI < 0.4), and extremely low (TCI < 0.2). Taking into account both slope and aspect change rates, the formula is:

[0095]

[0096] TRI is the topographic relief index, TPI is the topographic location index, Sstd is the standard deviation of slope, and C is the curvature; α, β, γ, and δ are the contributions of each topographic factor, calculated as follows:

[0097] First, construct the terrain factor matrix T = [TRI, TPI, Sstd, C]; second, obtain the contribution of each factor through PCA analysis:

[0098] In complex terrain regions (mountains, canyons, etc.), the ESRGAN model has relatively high initial weights due to its strong ability to reconstruct detailed textures. In gently sloping terrain regions (plains, farmland, etc.), the EDSR and VDSR models have high initial weights because they perform well in preserving large-scale structures. In moderately complex transitional terrain regions (hills, plateaus, shallow valleys, etc.), the RCAN model has high initial weights because its residual channel attention mechanism can adaptively select and strengthen important features, making it suitable for handling transitional regions and boundary areas with large variations in terrain complexity. The initial weight allocation follows normalization constraints to ensure that the sum of all model weights equals 1. A continuous function form is used to achieve a smooth transition of weights, avoiding spatial discontinuities that may be caused by hard thresholding, and ensuring the rationality of weight allocation at the boundaries of different terrain types and the consistency of super-resolution reconstruction results.

[0099] S50, based on the weights of the super-resolution sub-models, weights and fuses the reconstructed dataset output by the super-resolution combined model to generate the final high-resolution solar radiation data, which can maintain the hourly temporal resolution of the original data.

[0100] S60 acquires high-resolution solar radiation data and performs quality assessment and labeling on the generated high-resolution data. The final output data adopts standard geospatial data formats: GeoTIFF (suitable for GIS system integration); NetCDF (suitable for scientific computing and analysis); HDF5 (suitable for big data storage and processing). High-resolution solar radiation data includes: spatial resolution: 1 km; temporal resolution: hourly; coordinate system: WGS84; projection method: equal latitude and longitude projection; data unit: W / m². 2 Quality control code: 0-2 Level 3 quality grade.

[0101] In this embodiment of the invention, Figure 7 The diagram illustrates the process of quality assessment and labeling of the generated high-resolution data. A three-layer quality control system is adopted when assessing and labeling the generated high-resolution data: the first layer is numerical rationality check to ensure that the radiation value is within the physically possible range; the second layer is temporal continuity check to identify abnormal abrupt changes; and the third layer is spatial consistency check to ensure reasonable transition between adjacent pixels.

[0102] In the numerical rationality check, the physical range constraints are radiation value, temperature, and reflectivity, and unreasonable data are marked as outliers. In the temporal continuity check, time series analysis is used, and time anomaly detection is performed through sliding window to mark time anomalies. In the spatial consistency check, spatial correlation analysis is performed: neighborhood window detection is used to mark spatial anomalies.

[0103] Quality assessment is conducted by comparing data with ground observation data: calculating statistical indicators such as root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient (R); analyzing accuracy performance under different seasons and terrain conditions; and evaluating data quality during extreme weather events.

[0104] Then, based on the quality assessment results, a quality control code is assigned to each data point: 0 indicates the data is correct and usable, 1 indicates the data is questionable and should be used with caution, and 2 indicates the data is incorrect and its use is not recommended. High-resolution data, combining the quality control code and the overall quality score, is output along with the high-resolution data and compared with the site observation data to provide users with a data quality reference.

[0105] like Figure 6 The diagram illustrates the super-resolution reconstruction process of a radiation dataset using a super-resolution combined model. The super-resolution reconstruction process is described as follows:

[0106]

[0107] Among them, f i Let i represent the i-th super-resolution sub-model. It is the 1km simulation result output by the i-th super-resolution sub-model, that is, the reconstructed dataset result output by the i-th super-resolution sub-model;

[0108] The reconstructed dataset output by the super-resolution combined model is weighted and fused based on the weights of the super-resolution sub-models. The weighted fusion process is represented as follows:

[0109]

[0110] Among them, w i These are the weight coefficients of the i-th super-resolution sub-model, where n is the total number of super-resolution sub-models. For the weighted fusion output representation, the 1km resolution reconstruction results of the four super-resolution models are linearly weighted and fused according to the optimized weights during weighted fusion. The fusion process is performed pixel by pixel. For each 1km grid point, the weighted average of the output values ​​of the four models is calculated, and the weight coefficients are dynamically adjusted according to the terrain features of that point.

[0111] The fusion results maintain the hourly temporal resolution of the original data, generating a complete spatiotemporal continuous solar radiation dataset. The data format adopts the standard GeoTIFF format for easy subsequent application and analysis.

[0112] In this embodiment of the invention, when optimizing the weights of each super-resolution sub-model in the super-resolution ensemble model by minimizing the loss function, the weight optimization is achieved by minimizing the following loss function:

[0113]

[0114] Among them, I Obs For ground observation data, λ is the regularization parameter, and w terrain The weights are for terrain factors.

[0115] When optimizing the weights of each super-resolution sub-model in the super-resolution fusion model by minimizing the loss function, measured data from ground radiation observation stations within the study area are used as a validation benchmark. The loss function consists of two parts: the main term is the mean square error between the model fusion result and the ground observation data, used to ensure reconstruction accuracy; the constraint term is the deviation between the current weights and the terrain reference weights, used to maintain terrain adaptability.

[0116] The optimization process employs the gradient descent algorithm, iteratively updating the weight parameters until the loss function converges. The regularization parameter λ is set to 0.01 to ensure accuracy while avoiding overfitting. Ultimately, the optimal weight combination adapted to the current geographical environment is obtained.

[0117] In this embodiment of the invention, the output results of multiple models are weighted and fused, and the weights of the super-resolution sub-model are optimized by combining high-resolution terrain factors. This achieves spatial downscaling from 5km to 1km resolution. While maintaining the hourly temporal resolution of the original data, the spatial accuracy of solar radiation data is significantly improved. This method can capture more surface heterogeneity information while meeting the application requirements of high temporal resolution. It provides high-quality data support for climate research, solar energy resource assessment, precision agriculture and other fields. The method is also suitable for batch processing of large-scale areas and supports parallel computing.

[0118] This invention provides a method for preprocessing downwave radiation data from the Earth's surface. Figure 2 This diagram illustrates the implementation flow of a method for preprocessing downwave radiometric data from the Earth's surface. The method specifically includes:

[0119] S101: Acquire surface downwave radiation data, use the three-standard-deviation principle combined with an adaptive threshold to detect outliers in the surface downwave radiation data, mark outliers in the surface downwave radiation data, and output surface downwave radiation data after removing outliers.

[0120] This embodiment uses the entire territory of China as the study area and 2008 as the typical case year. The surface downwave radiation data is the 5km resolution surface downwave radiation data of China in 2008. The surface downwave radiation data comes from the National Tibetan Plateau Scientific Data Center (https: / / data.tpdc.ac.cn / en / data / 00c0a388-2b53-4592-b6c9-5011532478c5), with a temporal resolution of hourly and a spatial coverage of 71.025°E-141.025°E and 14.975°N-59.975°N. Figure 2 The image shows a raw data map of downward shortwave radiation data at the Earth's surface.

[0121] S102, Load the surface downwave radiation data after removing outliers, analyze the time series characteristics of the surface downwave radiation data, identify and remove data points affected by clouds, and obtain the surface downwave radiation data after cloud cover removal.

[0122] S103 acquires surface downwave radiometric data after cloud masking removal. The downwave radiometric data is standardized using the maximum-minimum method, normalizing the data to the 0-1 range to ensure comparability across different periods. Format conversion transforms multiple data sources into a standard geospatial data format, and spatiotemporal scale matching is performed on the surface downwave radiometric data to output a standardized radiometric dataset. Outlier detection and cloud masking removal remove redundant and invalid data, reducing the amount of data required for subsequent processing and analysis. This not only improves data processing efficiency but also reduces computational costs and storage requirements, making data processing and analysis more efficient and economical.

[0123] In this embodiment of the invention, during the preprocessing of surface downwave radiation data, a three-standard-deviation principle combined with an adaptive threshold is used to detect outliers in the surface downwave radiation data. This effectively identifies and removes outliers from the data. Furthermore, the dynamic threshold method can adapt to the characteristics of different climatic sub-regions, ensuring the accuracy and adaptability of outlier detection. The data after outlier removal is more accurate and reliable, reducing errors and biases caused by outliers. Identifying and removing data points affected by clouds effectively reduces the interference of clouds on radiation data, overcoming the problem of abnormally low radiation values ​​caused by the significant shading effect of clouds on solar radiation. By removing these cloud-affected data points, data that more accurately reflects the surface downwave radiation situation can be obtained, improving the accuracy and usability of the data.

[0124] This invention provides a method for outlier detection in surface downwave radiation data using a three-standard-deviation principle combined with an adaptive threshold. Figure 3 The diagram illustrates the implementation process of a method for outlier detection in surface downdraft shortwave radiation data using a three-standard-deviation principle combined with an adaptive threshold. Specifically, this method includes:

[0125] S1011, Load surface downwave radiation data, identify the climate sub-regions corresponding to the surface downwave radiation data, wherein the climate sub-regions include, but are not limited to, plateaus, plains, coasts, and lakes, and obtain at least one set of sub-region radiation datasets;

[0126] S1012, a sliding spatiotemporal window is used to perform spatiotemporal window segmentation on the sub-region radiation dataset to obtain at least one set of dynamic spatiotemporal window segments. The mean and standard deviation of the dynamic spatiotemporal window segments in the sub-region radiation dataset are calculated, and the initial anomaly threshold range of the dynamic spatiotemporal window segments is set using the three-times-standard-deviation principle.

[0127] The initial anomaly threshold range is represented as follows:

[0128] α0=μ i ±3σ i

[0129] Where α0 represents the initial anomaly threshold interval, μ i ,σ i These are the mean and standard deviation of the dynamic spatiotemporal window segment, respectively.

[0130] The size of the sliding spacetime window is represented as follows:

[0131] W c =W base ×(1+q W ·C i +κ·Ai )

[0132] Among them, W c W base These represent the size of the sliding spacetime window and the base window size, respectively. W κ represents the weighting coefficients of the sliding spatiotemporal window and the terrain factor weighting coefficient, respectively, and C i A i These are the slope component and the aspect effect coefficient, respectively.

[0133] S1013, identify atmospheric transmittance and solar zenith angle within a dynamic spatiotemporal window segment, and adjust the initial anomaly threshold range based on atmospheric transmittance and solar zenith angle to obtain the adjusted adaptive threshold.

[0134] The adaptive threshold is expressed as:

[0135]

[0136] Among them, W α For physical adjustment of thresholds, T0,θ,T s These are the solar coefficient, solar zenith angle, and atmospheric transmittance, respectively.

[0137] S1014, based on the adjusted adaptive threshold, performs outlier detection on the surface downwave radiation data within the dynamic spatiotemporal window segment, marks outliers in the surface downwave radiation data, and outputs the surface downwave radiation data after removing outliers.

[0138] In this embodiment of the invention, when using the three-standard-deviation principle combined with an adaptive threshold to detect outliers in surface downwave radiation data, a sliding spatiotemporal window is used to segment the sub-region radiation dataset into spatiotemporal windows. The initial outlier threshold interval is adjusted based on atmospheric transmittance and solar zenith angle to obtain an adjusted adaptive threshold. This significantly improves the accuracy and reliability of outlier detection in surface downwave radiation data. Furthermore, by integrating statistical laws and atmospheric physical mechanisms, it overcomes the limitations of traditional methods in areas with complex terrain and special weather conditions. The adaptive threshold adjustment mechanism allows for dynamic adjustment of the outlier threshold based on factors such as real-time atmospheric transmittance and solar zenith angle, without the need to pre-set a fixed threshold. This dynamic threshold adjustment capability enables the method to better cope with various complex environmental conditions and data changes, enhancing the method's adaptability and flexibility.

[0139] This invention provides a method for identifying and removing data points affected by cloud cover. Figure 4 This diagram illustrates the implementation flow of a method for identifying and removing data points affected by clouds. The method specifically includes:

[0140] S1021: Obtain surface downwave radiation data after removing outliers. Based on a spatiotemporal context detection algorithm, a cloud contamination threshold is preset. By analyzing the time-series characteristics of the surface downwave radiation data, the actual clear-sky radiation value is determined, and the radiation characteristic ratio is calculated in conjunction with the theoretical clear-sky radiation value. When the radiation characteristic ratio exceeds the preset cloud contamination threshold, the data point is determined to have cloud contamination. This effectively distinguishes between normal data points and data points affected by clouds, avoiding misjudgments and omissions, thereby improving data accuracy.

[0141] S1022, Analyze the time series characteristics of surface downwave radiation data, and determine the actual value of clear-sky radiation based on the time series characteristics of surface downwave radiation data;

[0142] S1023, determine the radiation characteristic ratio by combining the actual value of clear-sky radiation and the theoretical value of clear-sky radiation;

[0143] S1024, Determine whether the radiation characteristic ratio corresponding to the downwave radiation data on the ground surface exceeds the preset cloud pollution judgment threshold.

[0144] The formula for calculating the radiation characteristic ratio is as follows:

[0145]

[0146] in, D represents the ratio of radiation characteristics. sj D l These are the actual value of clear-sky radiation and the theoretical value of clear-sky radiation, respectively, T. ait Γ represents atmospheric transmittance and aerosol optical thickness, respectively.

[0147] S1025, If the radiation characteristic ratio corresponding to the surface downwave radiation data exceeds the preset cloud pollution judgment threshold, it is determined that the surface downwave radiation data has cloud pollution, and the surface downwave radiation data is marked as cloud pollution.

[0148] S1026. If the radiation characteristic ratio corresponding to the surface downwave radiation data does not exceed the preset cloud pollution judgment threshold, it is determined that the surface downwave radiation data does not have cloud pollution, and the surface downwave radiation data without cloud pollution is retained.

[0149] S1027, acquire surface downwave radiation data marked by cloud pollution, and calculate the optical thickness of the surface downwave radiation data;

[0150] The formula for calculating the optical thickness of surface downwave radiation data is as follows:

[0151]

[0152] Among them, Hef Optical thickness for downwave radiation data from the Earth's surface;

[0153] S1028, Determine whether the optical thickness of the downwave radiation data on the ground surface exceeds the preset thickness threshold.

[0154] If the thickness exceeds the preset thickness threshold, data points affected by clouds will be removed.

[0155] S1029, If the thickness does not exceed the preset thickness threshold, the spatiotemporal co-interpolation method is used to interpolate and reconstruct the surface downwave radiation data to obtain the reconstructed surface downwave radiation data. The spatiotemporal co-interpolation method can comprehensively consider the correlation of data in time and space, and use the information of surrounding normal data points to reconstruct data points affected by clouds. It can not only retain more data points to avoid excessive data loss, but also ensure the continuity and consistency of the reconstructed data in time and space, thereby enhancing the integrity of the data.

[0156] The integrated and reconstructed surface downwave radiation data, and the surface downwave radiation data without cloud pollution, are surface downwave radiation data after cloud masking removal.

[0157] In this embodiment of the invention, when identifying and removing data points affected by clouds, the spatiotemporal context detection algorithm and the preset cloud pollution judgment threshold are combined to accurately identify data points affected by clouds. At the same time, the spatiotemporal collaborative interpolation method can comprehensively consider the correlation of data in time and space, which can not only retain more data points and avoid excessive data loss, but also ensure the continuity and consistency of the reconstructed data in time and space, thereby enhancing the integrity of the data.

[0158] This invention provides a method for interpolating and reconstructing surface downwave radiation data using spatiotemporal collaborative interpolation. Figure 5 This document illustrates a flowchart of a method for interpolating and reconstructing surface downwave radiation data using spatiotemporal collaborative interpolation. The method specifically includes:

[0159] S1031, Load surface downwave shortwave radiation data, and construct a spatiotemporal reconstruction model for reconstructing surface downwave shortwave radiation data based on the surface downwave shortwave radiation data;

[0160] S1032, Perform topographic radiation correction and shadow correction on the surface downwave radiation data in the spatiotemporal reconstruction model, and output the spatiotemporal reconstruction model of the surface downwave radiation data after topographic radiation correction and shadow correction. In this embodiment, by performing topographic radiation correction and shadow correction on the surface downwave radiation data, the influence of topographic undulation on the radiation data can be eliminated, making the data more accurately reflect the actual radiation situation. Shadow correction can eliminate the interference of shadow areas on the radiation data, further improving the accuracy of the data.

[0161] S1033, solve for the data weighting coefficients of the surface downwave radiation data in the spatiotemporal reconstruction model, update the spatiotemporal reconstruction model with the weighting coefficients of the surface downwave radiation data, and obtain surface downwave radiation data based on spatiotemporal co-interpolation. The spatiotemporal co-interpolation method not only considers the temporal variation of the data, but also the spatial distribution characteristics of the data. It can generate data with high consistency in both time and space. Outputting spatiotemporally consistent data is of great significance for studying the spatiotemporal variation of surface downwave radiation, conducting regional climate analysis and environmental monitoring, and can improve the accuracy and reliability of research and application.

[0162] The spatiotemporal reconstruction model is represented as follows:

[0163]

[0164] in, These represent the surface downwave radiometric data based on spatiotemporal co-interpolation, the surface downwave radiometric data after topographic and shading correction, and the data for sample point m at time t, respectively. B represents the number of sample points m, and w represents the data for sample point m. m Here, θ represents the data weighting coefficients for the surface downwave shortwave radiation data, DSR(x,y,t) represents the raw surface downwave shortwave radiation data of sample point m at time t, and θ represents the data weighting coefficients for the surface downwave shortwave radiation data. p The angle between the slope normal and the direction of the sun; M(h) is the number of sample points m at a given spatiotemporal distance h. They are the spacetime positions x m ,(x m Topographically and shadow-corrected surface downwave radiometric data at +h).

[0165] In this embodiment of the invention, the super-resolution combined model is based on a parallel sub-module architecture. The parallel sub-module includes at least one set of super-resolution sub-models, such as ESRGAN, EDSR, VDSR, and RCAN models. The super-resolution combined model also includes an input layer, a denoising network, and an output layer. The input layer is connected to the denoising network, and the denoising network is connected to the parallel sub-module. The denoising network consists of an autoencoder and a decoder. A cross-model feature interaction layer and a multimodal joint representation layer are provided between the parallel sub-module and the output layer. The multimodal joint representation layer is a neural network model containing convolutional FConv layers. The noise reduction network uses a Butterworth filter to filter and denoise the radiation dataset, performs noise autoencoder recognition on the radiation dataset, and removes residual noise interference from the radiation dataset based on an improved residual neural network to complete the reconstruction of the radiation dataset. The super-resolution sub-model is used for super-resolution reconstruction of the radiation dataset. The maximum likelihood estimation method is introduced in the cross-model feature interaction layer. Based on the maximum likelihood estimation method, feature exchange nodes are defined when super-resolution reconstruction of the radiation dataset. The multimodal joint representation layer weights and fuses the reconstructed dataset output by the super-resolution combined model based on the weights of the super-resolution sub-model to generate the final high-resolution solar radiation data.

[0166] In this embodiment of the invention, the super-resolution sub-model includes the ESRGAN model, EDSR model, VDSR model, and RCAN model, and may also include the RDN model and CARN model. The ESRGAN model excels at processing images with rich texture details and is particularly suitable for radiometric data reconstruction in terrain-complex areas; the EDSR model performs excellently in maintaining overall structural consistency and is suitable for large areas with flat terrain; the VDSR model has a deep network structure and can learn complex nonlinear mapping relationships; the RCAN model integrates an attention mechanism, enabling it to adaptively focus on important features. Each model processes the input data at a 5km resolution, and through multi-layer feature extraction and reconstruction of the neural network model, generates the corresponding 1km resolution output result. The processing of the four models is independent and can be executed in parallel, significantly improving computational efficiency. Figure 6 As shown, the ESRGAN model, EDSR model, VDSR model, and RCAN model use a parallel data reading method for super-resolution reconstruction. The RCAN model includes an input layer, residual channel attention block, upsampling layer, and output layer connected in sequence. The VDSR model includes an input layer, deep convolutional planar network (CNN with more than or equal to 20 layers), and output layer connected in sequence. The EDSR model includes an input layer, multiple residual blocks, upsampling layer, and output layer connected in sequence. The ESRGAN model includes an input layer, convolutional layer, nested residual dense block, upsampling layer, and output layer connected in sequence.

[0167] Among them, methods for obtaining super-resolution ensemble models through pre-training on historical datasets include:

[0168] S201, Obtain historical dataset, insert preset square wave noise into historical dataset to obtain historical dataset with inserted noise, and divide historical dataset into training set and test set;

[0169] S202, Load the pre-built super-resolution combined model, freeze the parallel sub-modules, feature interaction layer and multimodal joint representation layer, and pre-train the denoising network using the training set to obtain the initial super-resolution combined model;

[0170] S203: Freeze the denoising network in the initial super-resolution combined model, use the training set to train the parallel sub-modules in the initial super-resolution combined model independently, and use the LAMB optimizer to optimize the hyperparameters of the parallel sub-modules during training, and output the independently trained super-resolution combined model.

[0171] S204. Initialize the feature interaction layer and preset the maximum likelihood node. Use the training set to jointly train the super-resolution combined model. Adjust the weights of the super-resolution sub-models during joint training to obtain the super-resolution combined model with adjusted weights.

[0172] S205: Obtain the test set, use the test set as input, execute the super-resolution combined model, output the test results, compare the test results with the ground observation data, and calculate the root mean square error, mean absolute error and correlation coefficient statistical indicators.

[0173] S206, determine whether the root mean square error, mean absolute error and correlation coefficient statistical indicators all meet the preset indicator thresholds;

[0174] Among them, whether the root mean square error, mean absolute error, and correlation coefficient statistical indicators all meet the preset threshold values ​​are as follows: Mean Absolute Error (MAE): <30W / m 2 Root mean square error (RMSE): <50W / m 2 Correlation coefficient (R): >0.85, processing time: <1 minute / image.

[0175] S207 If the root mean square error, mean absolute error, and correlation coefficient all meet the preset threshold values, output a converged super-resolution combined model.

[0176] S208 If any of the statistical indicators, such as root mean square error, mean absolute error, and correlation coefficient, does not meet the preset threshold, return to S202 and continue iterative training of the model.

[0177] After the above processing steps, the spatial resolution of solar radiation data across China was successfully improved from 5 km to 1 km, while the temporal resolution remained at the hourly level. Comparison with ground-based observation data shows a mean absolute error of 28.5 W / m². 2 The root mean square error is 45.2 W / m 2 The correlation coefficient reached 0.87, which is better than the processing results of traditional interpolation methods and single super-resolution models.

[0178] In terms of data processing efficiency, the processing time for a single hour of data is less than one minute, meeting the timeliness requirements of operational applications. The generated high-resolution data can clearly show the impact of terrain undulations on solar radiation distribution, providing high-quality data support for applications such as solar energy resource assessment and agricultural meteorological services.

[0179] The application scenarios of this application are as follows: climate change research: providing high-precision solar radiation data to support climate model research; solar energy resource assessment: providing a data foundation for solar power plant site selection and power generation prediction; precision agriculture: supporting crop growth models and irrigation decision systems; urban planning: analyzing urban heat island effect and building energy consumption; ecological environment monitoring: assessing vegetation photosynthesis and ecosystem health.

[0180] In summary, this invention provides a spatial downscaling method for surface downwave radiation based on super-resolution reconstruction technology. In this embodiment, the outputs of multiple models are weighted and fused, and the weights of the super-resolution sub-models are optimized using high-resolution topographic factors. This achieves spatial downscaling from 5km to 1km resolution, significantly improving the spatial accuracy of solar radiation data while maintaining the hourly temporal resolution of the original data. It captures more surface heterogeneity information and meets the requirements of high temporal resolution applications, providing high-quality data support for climate research, solar energy resource assessment, precision agriculture, and other fields. Furthermore, this invention supports automatic identification and parsing of various data source formats, including common formats such as NetCDF, HDF5, and GeoTIFF. The quality control algorithm automatically identifies and handles outliers, missing values, and cloud-contaminated pixels. Data standardization ensures consistency and comparability of data from different sources.

[0181] It should be noted that, for the sake of simplicity, the foregoing embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to the present invention. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0182] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still combine, add, delete, or otherwise adjust the features of the various embodiments of the present invention according to the circumstances without conflict or creative effort, thereby obtaining different technical solutions that do not fundamentally depart from the concept of the present invention. These technical solutions also fall within the scope of protection of the present invention.

Claims

1. A spatial downscaling method for surface downwave shortwave radiation based on super-resolution reconstruction technology, characterized in that, include: Acquire low-resolution surface downwave radiation data, preprocess the surface downwave radiation data, and output a standardized radiation dataset; Collect historical datasets and pre-train a super-resolution ensemble model using these datasets. The radiation dataset is loaded in real time, and a super-resolution ensemble model is used to perform super-resolution reconstruction of the radiation dataset to obtain at least one set of high-resolution reconstructed datasets. Obtain the reconstructed dataset, identify the terrain factor information in the reconstructed dataset, and combine the ground observation station data and terrain factor information to optimize the weights of each super-resolution sub-model in the super-resolution combined model by minimizing the loss function; The reconstructed dataset output by the super-resolution combined model is weighted and fused based on the weights of the super-resolution sub-models to generate the final high-resolution solar radiation data.

2. The surface downwave radiation spatial downscaling method based on super-resolution reconstruction technology as described in claim 1, characterized in that: The method for preprocessing surface downwave radiation data includes: Acquire surface downwave radiation data, use the three-standard-deviation principle combined with an adaptive threshold to detect outliers in the surface downwave radiation data, mark outliers in the surface downwave radiation data, and output surface downwave radiation data after removing outliers; Load the surface downwave shortwave radiation data after removing outliers, analyze the time series characteristics of the surface downwave shortwave radiation data, identify and remove data points affected by cloud cover, and obtain the surface downwave shortwave radiation data after cloud cover removal. Obtain surface downwave radiation data after cloud masking removal, standardize the surface downwave radiation data using the maximum-minimum method, and perform spatiotemporal scale matching processing on the surface downwave radiation data to output a standardized radiation dataset.

3. The surface downwave radiation spatial downscaling method based on super-resolution reconstruction technology as described in claim 2, characterized in that: The method for outlier detection in surface downwave radiation data using the three-standard-deviation principle combined with an adaptive threshold includes: Load surface downwave radiation data and identify the corresponding climate sub-regions; A sliding spatiotemporal window is used to perform spatiotemporal window segmentation on the sub-region radiation dataset to obtain at least one set of dynamic spatiotemporal window segments. The mean and standard deviation of the dynamic spatiotemporal window segments in the sub-region radiation dataset are calculated, and the initial anomaly threshold interval of the dynamic spatiotemporal window segments is set using the three-standard-deviation principle. Identify atmospheric transmittance and solar zenith angle within a dynamic spatiotemporal window segment, and adjust the initial anomaly threshold range based on atmospheric transmittance and solar zenith angle to obtain the adjusted adaptive threshold. Anomaly detection is performed on the surface downwave radiation data within the dynamic spatiotemporal window based on the adjusted adaptive threshold. Anomalies in the surface downwave radiation data are marked, and the surface downwave radiation data after removing anomalies is output.

4. The surface downwave radiation spatial downscaling method based on super-resolution reconstruction technology as described in claim 2, characterized in that: The method for identifying and removing data points affected by clouds includes: Obtain surface downwave radiation data after removing outliers, and preset cloud pollution judgment thresholds based on spatiotemporal context detection algorithms; Analyze the time series characteristics of surface downwave radiation data, and determine the actual clear-sky radiation value based on the time series characteristics of surface downwave radiation data; The radiation characteristic ratio is determined by combining the actual value of clear-sky radiation and the theoretical value of clear-sky radiation, and it is then used to determine whether the radiation characteristic ratio corresponding to the downwave radiation data at the ground surface exceeds the preset cloud pollution judgment threshold. If the radiation characteristic ratio corresponding to the downwave radiation data of the surface exceeds the preset cloud pollution judgment threshold, it is determined that there is cloud pollution in the downwave radiation data of the surface, and the downwave radiation data of the surface is marked as cloud pollution. If the radiation characteristic ratio corresponding to the surface downwave radiation data does not exceed the preset cloud pollution judgment threshold, it is determined that the surface downwave radiation data does not have cloud pollution, and the surface downwave radiation data without cloud pollution is retained. Acquire surface downwave radiation data marked by cloud pollution and calculate the optical thickness of the surface downwave radiation data; Determine whether the optical thickness of the downwave radiation data on the ground surface exceeds a preset thickness threshold. If the thickness exceeds the preset thickness threshold, remove data points affected by clouds. If the thickness does not exceed the preset thickness threshold, the spatiotemporal co-interpolation method is used to interpolate and reconstruct the surface downwave shortwave radiation data to obtain the reconstructed surface downwave shortwave radiation data. The integrated and reconstructed surface downwave radiation data, and the surface downwave radiation data without cloud pollution, are surface downwave radiation data after cloud masking removal.

5. The surface downwave radiation spatial downscaling method based on super-resolution reconstruction technology as described in claim 4, characterized in that: The method for interpolating and reconstructing surface downwave radiation data using spatiotemporal co-interpolation includes: Load surface downwave radiometric data and construct a spatiotemporal reconstruction model based on the surface downwave radiometric data to reconstruct the surface downwave radiometric data; Perform topographic and shadow corrections on the surface downwave radiometric data in the spatiotemporal reconstruction model, and output the spatiotemporal reconstruction model of the surface downwave radiometric data after topographic and shadow corrections; Solve for the data weighting coefficients of the surface downwave radiation data in the spatiotemporal reconstruction model, and update the spatiotemporal reconstruction model with the weighting coefficients of the surface downwave radiation data to obtain surface downwave radiation data based on spatiotemporal co-interpolation.

6. The surface downwave radiation spatial downscaling method based on super-resolution reconstruction technology as described in any one of claims 2-5, characterized in that: The method for obtaining a super-resolution ensemble model through pre-training on historical datasets includes: Obtain the historical dataset, insert the preset square wave noise into the historical dataset to obtain the historical dataset with inserted noise, and divide the historical dataset into training set and test set; Load the pre-built super-resolution ensemble model, freeze the parallel sub-modules, feature interaction layer and multimodal joint representation layer, and pre-train the denoising network using the training set to obtain the initial super-resolution ensemble model; Freeze the denoising network in the initial super-resolution combined model, use the training set to train the parallel sub-modules in the initial super-resolution combined model independently, and use the LAMB optimizer to optimize the hyperparameters of the parallel sub-modules during training, and output the independently trained super-resolution combined model. Initialize the feature interaction layer and preset the maximum likelihood node. Use the training set to jointly train the super-resolution combined model. Adjust the weights of the super-resolution sub-models during joint training to obtain the weight-adjusted super-resolution combined model. Obtain a test set, use the test set as input, execute the super-resolution combined model, output the test results, compare the test results with the ground observation data, calculate the root mean square error, mean absolute error and correlation coefficient statistical indicators, and determine whether the root mean square error, mean absolute error and correlation coefficient statistical indicators all meet the preset index thresholds. If the root mean square error, mean absolute error and correlation coefficient statistical indicators all meet the preset index thresholds, output the converged super-resolution combined model.

7. The surface downwave radiation spatial downscaling method based on super-resolution reconstruction technology as described in claim 6, characterized in that: The super-resolution combined model is based on a parallel sub-module architecture. The parallel sub-module includes at least one set of super-resolution sub-models, such as the ESRGAN model, EDSR model, VDSR model, and RCAN model. The super-resolution combined model also includes an input layer, a denoising network, and an output layer. The input layer is connected to the denoising network, and the denoising network is connected to the parallel sub-module. A cross-model feature interaction layer and a multimodal joint representation layer are set between the parallel sub-module and the output layer.

8. The surface downwave radiation spatial downscaling method based on super-resolution reconstruction technology as described in claim 7, characterized in that: When using a super-resolution combined model to perform super-resolution reconstruction of a radiation dataset, the super-resolution reconstruction process is represented as follows: Among them, f i Let i represent the i-th super-resolution sub-model. It is the 1km simulation result output by the i-th super-resolution sub-model; The reconstructed dataset output by the super-resolution combined model is weighted and fused based on the weights of the super-resolution sub-models. The weighted fusion process is represented as follows: Among them, w i is the weight coefficient of the i-th super-resolution sub-model, and n is the total number of super-resolution sub-models.

Images