Super-resolution method and system of key space weather parameters based on LIPF

Abstract

The present application provides a LIPF-based super-resolution method and system for key space weather parameters, which relates to the field of data analysis technology. Key space weather parameters are obtained based on a target area, and dual-version data are outputted respectively using an ionosphere-thermosphere numerical simulation model. A super-resolution model is built based on the LIIF algorithm, and supervised training and cross-validation are performed on the model using the dual-version data to obtain a converged super-resolution model. Finally, the hyper-parameters of the model are retrieved and optimized based on a genetic algorithm, which solves the technical problems in the prior art of image super-resolution algorithms when applied to the reconstruction of thermospheric atmospheric parameters, such as insufficient model generalization ability, high computational complexity, low practicality and efficiency, and poor adaptability. This achieves the technical effect of providing an efficient, accurate, and adaptable method for reconstructing thermospheric atmospheric parameters, which can effectively improve the resolution and accuracy of thermospheric atmospheric parameters, reduce computational complexity, and improve the practicality and efficiency of the algorithm.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and in particular to a method and system for super-resolution of key space weather parameters based on LIPF. Background Art

[0002] In the field of space weather, accurate acquisition of key parameters of the thermosphere is of great significance for weather forecasting, communication and navigation. However, due to the limitations of traditional observation methods, the obtained thermosphere parameter data often have low resolution and accuracy, which is difficult to meet actual needs. Therefore, studying how to improve the resolution and accuracy of key parameters of the thermosphere has become one of the current research hotspots in the field of space weather.

[0003] The existing solutions to the above problems are mainly image super-resolution algorithms based on deep learning technology. This algorithm maps the input low-resolution image to the high-resolution image space by training a deep neural network model, thereby achieving super-resolution reconstruction of the image. However, due to the particularity of space weather data, there may still be some limitations in the application of the above technology, such as insufficient model generalization ability and poor adaptability to specific weather conditions.

[0004] In summary, the image super-resolution algorithms existing in the prior art, when applied to the reconstruction of thermospheric atmospheric parameters, have the problems of insufficient model generalization ability, high computational complexity, low practicality and efficiency of the algorithms, and due to the complexity and diversity of space weather data, it is difficult for the existing algorithms to effectively adapt to the needs of thermospheric atmospheric parameter reconstruction under different weather conditions, resulting in technical problems with poor prediction effects. Summary of the invention

[0005] The present application provides a method and system for super-resolution of key space weather parameters based on LIPF, which is used to solve the technical problems that the image super-resolution algorithm in the prior art has insufficient model generalization ability, high computational complexity, low algorithm practicality and efficiency when applied to the reconstruction of thermospheric atmospheric parameters, and due to the complexity and diversity of space weather data, it is difficult for the existing algorithms to effectively adapt to the needs of thermospheric atmospheric parameter reconstruction under different weather conditions, resulting in poor prediction results.

[0006] In view of the above problems, the present application provides a method and system for super-resolution of key space weather parameters based on LIPF.

[0007] In a first aspect, the present application provides a method for super-resolution of key parameters of space weather based on LIPF, the method being applied to a super-resolution system of key parameters of space weather based on LIPF, the method comprising:

[0008] Determine a target area, collect key space weather parameters based on the target area, and obtain key space weather parameters to be processed;

[0009] Performing data processing according to the key space weather parameters to be processed to obtain the key space weather parameters, and using an ionosphere-thermosphere numerical simulation model to output dual-version data, respectively, wherein the dual-version data includes a high-precision data stream and a low-precision data stream;

[0010] A super-resolution model is constructed based on the LIIF algorithm, and supervised training and cross-validation are performed on the super-resolution model based on the dual-version data to obtain a converged super-resolution model;

[0011] Retrieve and optimize the hyperparameters of the super-resolution model based on genetic algorithms to obtain the best hyperparameter combination;

[0012] The converged super-resolution model is optimized by using the optimal super-parameter combination to obtain an optimized super-resolution model.

[0013] In a second aspect, the present application provides a space weather key parameter super-resolution system based on LIPF, the system comprising:

[0014] A parameter collection module is used to determine a target area, collect key space weather parameters based on the target area, and obtain key space weather parameters to be processed;

[0015] A data output module, configured to perform data processing according to the key space weather parameters to be processed to obtain the key space weather parameters, and output dual-version data using an ionosphere-thermosphere numerical simulation model, wherein the dual-version data includes a high-precision data stream and a low-precision data stream;

[0016] A model building module, used to build a super-resolution model based on the LIIF algorithm, and perform supervised training and cross-validation on the super-resolution model based on the dual-version data to obtain a converged super-resolution model;

[0017] The parameter optimization module is used to retrieve and optimize the hyperparameters of the super-resolution model based on the genetic algorithm to obtain the best hyperparameter combination;

[0018] The model optimization module is used to perform model optimization of the converged super-resolution model through the optimal super-parameter combination to obtain an optimized super-resolution model.

[0019] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0020] The LIPF-based space weather key parameter super-resolution method provided in the present application determines a target area, collects space weather key parameters based on the target area, and obtains the space weather key parameters to be processed; performs data processing according to the space weather key parameters to be processed to obtain the space weather key parameters, and uses an ionosphere-thermosphere numerical simulation model to output dual-version data respectively, wherein the dual-version data includes a high-precision data stream and a low-precision data stream; constructs a super-resolution model based on the LIIF algorithm, and performs supervised training and cross-validation on the super-resolution model based on the dual-version data to obtain a converged super-resolution model; retrieves and optimizes the hyperparameters of the super-resolution model based on a genetic algorithm to obtain the best hyperparameter combination; The model optimization of the converged super-resolution model is performed through the optimal hyperparameter combination to obtain an optimized super-resolution model, which solves the technical problems that the image super-resolution algorithm in the prior art has insufficient model generalization ability, high computational complexity, low algorithm practicability and efficiency when applied to the reconstruction of thermospheric atmospheric parameters, and due to the complexity and diversity of space weather data, it is difficult for the existing algorithm to effectively adapt to the needs of thermospheric atmospheric parameter reconstruction under different weather conditions, resulting in poor prediction effect. The method provides an efficient, accurate and adaptable thermospheric atmospheric parameter reconstruction method, which can effectively improve the resolution and accuracy of thermospheric atmospheric parameters, while reducing computational complexity and improving the practicability and efficiency of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A schematic diagram of the process of super-resolution of key parameters of space weather based on LIPF is provided for this application;

[0022] Figure 2 A schematic diagram of the structure of the space weather key parameter super-resolution system based on LIPF is provided for this application.

[0023] Explanation of the accompanying drawings: parameter acquisition module 11, data output module 12, model building module 13, parameter optimization module 14, model optimization module 15. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0025] Embodiment 1

[0026] like Figure 1 As shown, the present application provides a method for super-resolution of key parameters of space weather based on LIPF, the method is applied to a super-resolution system of key parameters of space weather based on LIPF, and the method includes:

[0027] Step S100: determining a target area, collecting key space weather parameters based on the target area, and obtaining key space weather parameters to be processed;

[0028] Specifically, first, it is necessary to clarify the specific needs and goals of collecting key space weather parameters, including specific geographical locations, altitude ranges, time windows, etc., use high-precision map data and historical space weather data to make more detailed divisions and selections of target areas, and finally determine the target area, that is, the target area refers to the local space area determined for space weather forecasting and analysis. Next, select appropriate collection equipment to collect key space weather parameters in the target area. The collection equipment can be satellites, ground observation stations, radars, etc. The collected data includes meteorological data such as temperature and air pressure, as well as space weather parameters such as ionosphere and magnetic field. The collected key space weather parameter data is transmitted to the data processing center in real time or regularly for subsequent processing. Among them, the collected data has different resolutions and coverage, providing comprehensive basic data for subsequent analysis.

[0029] Step S200: performing data processing according to the key space weather parameters to be processed to obtain the key space weather parameters, and using the ionosphere-thermosphere numerical simulation model to output dual-version data respectively, wherein the dual-version data includes a high-precision data stream and a low-precision data stream;

[0030] Step S300: constructing a super-resolution model based on the LIIF algorithm, and performing supervised training and cross-validation on the super-resolution model based on the dual-version data to obtain a converged super-resolution model;

[0031] Optionally, since the collected data is too complicated, the acquired key parameter data to be processed is preprocessed, that is, the collected raw data is cleaned, outliers, missing values ​​or erroneous data are removed, and the data is interpolated, smoothed or standardized as necessary to ensure the quality and consistency of the data. Then, the preprocessed key space weather parameters are used as input and imported into the ionosphere-thermosphere numerical simulation model to simulate the dynamic behavior of the ionosphere and thermosphere, and dual versions of data are output respectively, that is, under the same given space weather conditions, the ionosphere-thermosphere numerical simulation model TIEGCM is used to output high-precision and low-precision versions of data respectively. Among them, low-resolution data is used for super-resolution model training, while high-resolution data is used to verify whether the deep learning results based on LIIF algorithm super-resolution are correct. In addition, when obtaining low-resolution and high-resolution data simulated by TIEGCM, it is necessary to ensure that the data format is unified to facilitate the use of super-resolution models.

[0032] Among them, TIEGCM is a theoretical model developed by NCAR / HAO to explore the interaction between the ionosphere and the thermosphere. The model is a time-varying three-dimensional dynamic theoretical model, which is divided into 29 pressure layers in vertical resolution, with an altitude range of about 97 to 600 kilometers, and a horizontal resolution of 5°×5° in latitude and longitude. LIIF (Learning Continuous Image Representation with Local Implicit Image Function) is a model based on local implicit functions.

[0033] The high-precision data stream in the dual-version data obtained usually has higher resolution and accuracy, and can describe the structure and dynamic changes of the ionosphere and thermosphere in detail. The low-precision data stream may have lower resolution and accuracy, but can meet the needs of certain application scenarios, such as real-time warning, preliminary analysis, etc. By combining the two types of data to train and verify the super-resolution model, the accuracy of the model can be guaranteed.

[0034] Further, a deep learning model based on the LIIF algorithm is designed to determine the key parameters of the model, such as the network structure, loss function, and optimizer. Among them, a deep learning framework (such as TensorFlow, PyTorch, etc.) is used to define the network structure of the super-resolution model, including a feature extraction layer, an upsampling module, and the core part of the LIIF algorithm; a loss function suitable for the super-resolution task is defined, such as mean square error (MSE), mean absolute error (MAE), or structural similarity index (SSIM), etc., which will be used to measure the difference between the high-resolution image predicted by the model and the real high-resolution image. It can be understood that the LIPF (Local Implicit Physical Function) proposed in this application is a method for learning continuous physical field representation using local implicit image functions. This method is based on LIIF and aims to map each pixel in the input image to a continuous representation space by learning a local implicit function, thereby achieving efficient representation and processing of the image. Since LIIF is a model based on local implicit functions, it is necessary to design a suitable network structure to learn local implicit functions to complete the construction of the super-resolution model.

[0035] Specifically, the general steps of LIPF are: first, a neural network model is used to learn a local implicit function. This function takes the coordinates of each pixel in the input image as input and outputs a continuous representation vector. This representation vector can be regarded as the position of the pixel in the representation space to complete the learning of the local implicit function; next, for each pixel in the input image, the local implicit function obtained by learning is mapped to a continuous representation space. This process can be regarded as re-representing the input image, converting discrete pixel values ​​into continuous representations for image mapping; then, in the representation space, various image processing operations are performed on the continuous representation, such as interpolation, transformation, filtering, etc. Since the representation space is continuous, these operations can be more flexibly applied to image data and usually have higher accuracy and efficiency; after completing the processing of the continuous representation, the result is mapped back to the original image space. This process converts the continuous representation back into discrete pixel values ​​for further analysis or application. The basic steps of the LIIF method are the same when processing high-resolution and low-resolution data, but there may be some differences: for high-resolution data, more complex models and larger representation spaces are usually required to capture richer image information, and more training data and deeper neural network structures may be required when learning local implicit functions; for low-resolution data, simpler models and smaller representation spaces can be used to improve computational efficiency and reduce storage space, and when learning local implicit functions, the complexity and number of parameters of the neural network can be reduced to adapt to the characteristics of low-resolution images. When processing high / low resolution data, the complexity of the model and representation space can be adjusted according to specific needs to achieve better results.

[0036] Furthermore, when the model framework is fully built, it is necessary to use low-resolution simulated data and high-resolution simulated data as input and output respectively, construct a training data set, and perform cross-validation to more accurately evaluate the performance of the model and reduce the risk of overfitting. Specifically, the dual-version data is divided into a training set, a validation set, and a test set. A batch of low-resolution images and high-resolution images are obtained from the training set. The low-resolution images are input into the super-resolution model to obtain the predicted high-resolution images. The loss between the predicted image and the high-resolution image is calculated based on the loss function. The optimizer is used to update the parameters of the model to minimize the loss. After each iteration, the validation set is used to evaluate the performance of the model and the performance indicators are recorded. If the performance of the validation set no longer improves after a certain iteration, an early stopping strategy is used to prevent overfitting.

[0037] In order to obtain a converged super-resolution model, it is also necessary to use cross-validation technology to evaluate the performance and stability of the model. Specifically, the data set is divided into K folds (such as K = 5), K-1 folds are used as training sets, and the remaining 1 fold is used as a validation set. Supervised training is performed on the training set and validation set, and the performance indicators on the validation set are recorded. The above steps are repeated K times, and a different fold is selected as the validation set each time. The average and standard deviation of the performance indicators of K validations are calculated to evaluate the performance and stability of the model.

[0038] Next, based on the results of supervised training and cross-validation, the model with the best performance is selected as the converged super-resolution model, which is usually determined by comparing the performance indicators of the validation set or cross-validation. Finally, the converged super-resolution model is finally evaluated using the test set, and the model is adjusted and optimized as needed.

[0039] By building a super-resolution model based on the LIIF algorithm and performing supervised training and cross-validation on the super-resolution model based on the dual-version data, a converged super-resolution model is obtained, and continuous representation of the image is achieved, thereby providing a flexible, efficient and powerful method for image processing, completing parameter reconstruction, and improving the resolution and accuracy of thermospheric atmospheric parameters, while reducing computational complexity and improving the practicability and efficiency of the algorithm.

[0040] Step S400: searching and optimizing the hyperparameters of the super-resolution model based on a genetic algorithm to obtain an optimal hyperparameter combination;

[0041] Step S500: Optimizing the converged super-resolution model by using the optimal super-parameter combination to obtain an optimized super-resolution model.

[0042] Exemplarily, define the search space of hyperparameters, such as learning rate, batch size, number of network layers, activation function type, etc., initialize a population, and each individual in the population represents a set of hyperparameter combinations. Define a fitness function to evaluate the quality of each individual. The fitness function can be a performance metric on the validation set of the super-resolution model, and calculate the fitness value of each individual in the population. Use a certain selection strategy to select excellent individuals from the current population for subsequent genetic operations. The strategy can be set by yourself and is not limited here. Randomly select two parent individuals for cross-pairing, use a certain crossover strategy (such as single-point crossover, multi-point crossover, uniform crossover, etc.) to generate new offspring individuals, and mutate the hyperparameters in the offspring individuals with a certain mutation probability. This mutation operation can increase the diversity of the population and prevent the algorithm from converging to the local optimal solution too early. Add the new individuals generated after selection, crossover and mutation to the new population, and keep the population size unchanged. In this process, some individuals with low fitness may need to be eliminated, and check whether the termination conditions are met (such as reaching the maximum number of iterations, the fitness value has not improved significantly within a certain number of iterations, etc.). If the termination conditions are met, the optimal individual (that is, the best hyperparameter combination) is output, otherwise, return to calculate the fitness value of each individual and continue to iterate. Use the best hyperparameter combination found by the genetic algorithm to train the super-resolution model, and evaluate the performance of the model on the test set. Finally, according to the best hyperparameter combination obtained by the genetic algorithm, update the configuration of the super-resolution model, and use the new best hyperparameter combination to retrain the model, which may involve multiple training cycles. After the training is completed, select the best model weights based on the performance on the validation set to obtain the optimized super-resolution model.

[0043] By adjusting and optimizing the super-resolution model based on the genetic algorithm, the model can automatically find the hyperparameter combination that best suits the current data set and model structure, thereby significantly improving the performance of the model, more accurately restoring the details of high-resolution images, and improving the visual quality of the image. The genetic algorithm can quickly locate the optimal range of hyperparameter combinations, avoiding wasting computing resources in invalid parameter spaces, and using the best hyperparameter combination for model training can speed up the convergence of the model and reduce training time. It also improves the generalization ability and adaptability of the model. Ultimately, it effectively improves the resolution and accuracy of the thermospheric atmospheric parameters, while reducing computational complexity and improving the practicality and efficiency of the algorithm.

[0044] Further, data processing is performed according to the key space weather parameters to be processed to obtain the key space weather parameters, and the ionosphere-thermosphere numerical simulation model is used to output dual-version data respectively, and the dual-version data includes a high-precision data stream and a low-precision data stream. Step S200 of the present application also includes:

[0045] Step S210: performing data cleaning and data preprocessing on the key space weather parameters to be processed to obtain preprocessing key parameters;

[0046] Step S220: performing data classification and marking on the preprocessing key parameters, performing feature extraction, labeling the preprocessing key parameters, and obtaining labeled key parameters;

[0047] Step S230: correcting and eliminating the labeled key parameters to obtain space weather key parameters;

[0048] Step S240: inputting the key space weather parameters into the ionosphere-thermosphere numerical simulation model and outputting dual-version data.

[0049] Specifically, the key space weather parameters include but are not limited to parameters such as thermospheric temperature, density, and ionospheric electron density. The above parameters are cleaned to remove or correct outliers, missing values, duplicate values, or erroneous data in the key space weather parameters to be processed, and the cleaned data is standardized and normalized for subsequent analysis and training. After cleaning and preprocessing, the data is converted into a format suitable for subsequent analysis, i.e., preprocessing key parameters. Further, according to the nature or characteristics of the key space weather parameters, they are divided into different categories or subsets, and the most representative features or attributes are extracted, which can accurately reflect the key information of space weather. Subsequently, labels or identifiers are assigned to the extracted features, and labeled key parameters are the basis for model training and evaluation. In the labeling process, it is necessary to further correct or eliminate some data that do not meet the requirements or are of low quality. The key space weather parameters obtained after correction and elimination are more accurate and reliable, and are used for subsequent ionosphere-thermosphere numerical simulation. Finally, the acquired key space weather parameters are used as the input of the ionosphere-thermosphere numerical simulation model. Based on the input key space weather parameters, the model outputs dual-version data of the ionosphere-thermosphere through numerical calculation and simulation, including high-precision and low-precision simulation results of physical quantities such as ionospheric density, temperature, and electric field.

[0050] Through the steps of data cleaning, preprocessing, classification and labeling, the quality and accuracy of key space weather parameters are improved, and they are applied to the ionosphere-thermosphere numerical simulation model to obtain more accurate simulation results and forecast information.

[0051] Furthermore, a super-resolution model is constructed based on the LIIF algorithm, and supervised training and cross-validation are performed on the super-resolution model based on the dual-version data to obtain a converged super-resolution model. Step S300 of the present application also includes:

[0052] Step S310: Select the LIIF algorithm and design a network structure of a super-resolution model based on the LIIF algorithm, including an encoder and a decoder;

[0053] Step S320: determining a loss function and selecting an optimizer to update the parameters of the super-resolution model;

[0054] Step S330: After the parameters are updated, a super-resolution model is obtained.

[0055] Optionally, based on the LIIF algorithm, a network structure capable of learning mapping from low-resolution images to high-resolution images is constructed, including an encoder and a decoder. The encoder is usually used to extract features from the input low-resolution image, and a convolutional neural network (CNN) can be used as an encoder, which contains multiple convolutional layers, activation functions and possible pooling layers or stride convolutions to reduce the spatial size of the feature map. Specifically, the input layer is used to receive the low-resolution image as input, and multiple convolutional layers are used to extract image features. The size and number of convolution kernels of each layer can be adjusted as needed. A nonlinear activation function is applied after each convolutional layer. Finally, a pooling layer or stride convolution is used to reduce the spatial size of the feature map so as to capture more global information at a higher level. The decoder is used to reconstruct a high-resolution image from the features extracted by the encoder. Since the LIIF algorithm is a coordinate-based method, the decoder will output a coordinate-related feature vector, which will be used to predict high-resolution pixel values ​​at any coordinate. Specifically, feature mapping is first performed to convert the output feature map of the encoder into a form suitable for use by the LIIF algorithm; next, the LIIF layer is used to predict high-resolution pixel values ​​from the input low-resolution coordinates and features. The LIIF layer is the core part of the model. It accepts low-resolution coordinates and feature mapping as input and outputs pixel value predictions at high-resolution coordinates. The LIIF layer usually includes a multi-layer perceptron; since LIIF is a coordinate-based prediction method, it does not require a traditional upsampling layer. However, in order to convert the prediction result into an image that matches the target resolution, some post-processing steps such as interpolation or averaging operations may be required. Finally, the network structure is built.

[0056] Further, the loss function is determined. For super-resolution tasks, commonly used loss functions include mean square error loss, mean absolute error loss, and structural similarity index loss, etc. Those skilled in the art can select or combine these loss functions as needed. Since LIIF is a coordinate-based prediction method, the loss function should be able to measure the difference between the predicted pixel value and the actual pixel value at any coordinate. Next, the optimizer is used to update the parameters of the model according to the gradient of the loss function. Commonly used optimizers include stochastic gradient descent (SGD), Adam, RMSprop, etc., and the appropriate optimizer and its hyperparameters (such as learning rate, momentum, etc.) can be selected according to the complexity of the model and the characteristics of the training data. When the parameters are updated, the super-resolution model is obtained.

[0057] Furthermore, supervised training and cross-validation are performed on the super-resolution model based on the dual-version data to obtain a converged super-resolution model. Step S300 of the present application further includes:

[0058] Step S340: selecting excellent samples from the dual-version data as a sample data set;

[0059] Step S350: obtaining a preset data division ratio, and dividing the sample data set into a training data set and a verification data set according to the preset data division ratio;

[0060] Step S360: performing supervised training on the super-resolution model using the training data set, and when the model output result tends to converge, verifying the output result of the super-resolution model using the verification data set;

[0061] Step S370: obtaining a preset model verification accuracy index, and when the model accuracy of the super-resolution model meets the preset model verification accuracy index, obtaining a converged super-resolution model.

[0062] Exemplarily, supervised training and cross-validation are performed on the obtained super-resolution model based on the dual-version data to converge the super-resolution model. First, excellent samples are selected from the dual-version data as sample data sets based on indicators such as data clarity, detail richness, and noise level. The ratio of the training data set and the verification data set is set in advance. Common division ratios include 80% training set and 20% verification set. According to the preset data division ratio, the sample data set is randomly divided into a training data set and a verification data set, and the randomness and uniformity of the data during the division process are ensured. Load the previously constructed super-resolution model based on the LIIF algorithm, initialize the parameters of the model, and use the training data set to perform multiple rounds of supervised training on the super-resolution model. In each training batch, low-precision data is input into the model, and the loss function value is calculated by forward propagation. The optimizer is used to update the weight of the model to minimize the loss function value. Repeat the training process until the performance of the model on the verification set is stable or the preset stop condition is met. When the model output results tend to converge, the output results of the super-resolution model are verified using the validation data set, and the low-resolution images in the validation data set are input into the model to obtain the prediction results of the high-resolution images, and the difference or similarity between the prediction results and the actual high-resolution images is calculated. The performance of the super-resolution model on the validation data set is evaluated according to the preset model verification accuracy index. If the accuracy of the model meets the preset index, it is considered that the model has achieved a good performance level in the current training stage, and the parameters and structural information of the current model are saved to obtain a converged super-resolution model. Among them, the converged super-resolution model still has room for optimization, and can be further fine-tuned in subsequent tasks, adjusting the parameters, structure or training strategy of the model, etc., and repeating the above training, verification and saving process to obtain a super-resolution model with better performance.

[0063] Furthermore, the present application further includes step S380, and step S380 further includes:

[0064] Step S381: Divide the training data set into multiple non-overlapping sub-training sets;

[0065] Step S382: arbitrarily selecting the first sub-training set as a sub-validation data set, and the remaining sub-training sets as sub-training data sets;

[0066] Step S383: cross-training the first converged super-resolution model according to the sub-validation data set and the sub-training data set to obtain a first performance indicator;

[0067] Step S384: performing an iterative operation on the sub-training set, obtaining the performance index corresponding to the sub-training set and performing mean calculation to obtain the overall performance result of the first convergent super-resolution model;

[0068] Step S385: Analyze based on the overall performance results to obtain a converged super-resolution model.

[0069] Specifically, when using cross-validation to evaluate and optimize a super-resolution model that has converged once, the training data set is first evenly divided into K non-overlapping sub-training sets, where K is the number of cross-validation folds, and the first sub-training set is randomly selected from the K sub-training sets as the sub-validation data set, and the remaining K-1 sub-training sets are used as sub-training data sets. During the model training process, the super-resolution model that has converged once is trained using the selected K-1 sub-training data sets, and the trained model is validated using the selected sub-validation data set, and the corresponding performance indicators, i.e., the first performance indicators, are recorded. The process of selecting sub-validation sets and cross-training is repeated, and different sub-training sets are selected as sub-validation data sets each time until all sub-training sets have been used as sub-validation data sets, and the performance indicators obtained from each iteration are recorded. The performance indicators obtained from the K iterations are then summarized, and the average value of the K performance indicators is calculated. This average value represents the overall performance result of the super-resolution model that has converged once. Based on the overall performance results, the generalization ability and stability of the super-resolution model that converges once can be evaluated. If the overall performance results meet the preset requirements, then the super-resolution model that converges once can be considered effective and can be used as the final converged super-resolution model. If it does not meet the requirements, it may be necessary to adjust the model structure, parameters or training strategy, and repeat the entire training and validation process. In this way, cross-validation can more accurately evaluate the performance of the model, reduce the risk of overfitting, and select a more robust model.

[0070] Furthermore, the hyperparameters of the super-resolution model are retrieved and optimized based on the genetic algorithm to obtain the best hyperparameter combination. Step S400 of the present application also includes:

[0071] Step S410: setting individuals and the number of individuals, and randomly generating a set of hyperparameters for the individuals as an initial solution;

[0072] Step S420: using the super-resolution model and the training data set, training and verifying the hyper-parameter combination to obtain individual fitness;

[0073] Step S430: sorting based on the individual fitness, setting a fitness threshold, and selecting the individual whose individual fitness meets the fitness threshold as the parent individual;

[0074] Step S440: performing a crossover operation on the parent individuals to generate new offspring individuals, performing a mutation operation based on the new offspring individuals, combining the new offspring individuals with the parent individuals into a new hyperparameter combination, and cutting the new hyperparameter combination according to the fitness threshold, selecting the individuals with high fitness to retain;

[0075] Step S450: Preset a termination condition. When the termination condition is met, stop the iteration and output the optimal individual as the optimal hyperparameter combination.

[0076] Exemplarily, a genetic algorithm is used to automatically optimize the hyperparameters of the super-resolution model to obtain better model performance. The process of the genetic algorithm is: first determine the number of individuals in the genetic algorithm, that is, the population size, and each individual represents a hyperparameter combination. A set of hyperparameters is randomly generated for each individual as the initial solution, and the hyperparameter combination of each individual is trained and verified using the super-resolution model and the training data set, and the performance index on the verification set is used as the fitness value of the individual. Then, the individuals are sorted according to their fitness values, and a fitness threshold is set. Individuals with fitness values ​​higher than the threshold are selected as parent individuals, and two parent individuals are randomly selected. New offspring individuals are generated through crossover operations (such as single-point crossover, multi-point crossover, etc.), where you that the crossover operation is to exchange some gene fragments of two parent individuals to generate new gene combinations. The newly generated offspring individuals are mutated to introduce new gene combinations. The mutation operation refers to randomly changing one or some gene values ​​of the offspring individuals. The mutated new offspring individuals are combined with the parent individuals into a new population for the next round of iteration. In the new population, pruning is performed according to the fitness threshold, only individuals with higher fitness values ​​are retained, and individuals with lower fitness values ​​are eliminated, which helps to maintain the diversity and optimization ability of the population. One or more termination conditions are preset, such as reaching the maximum number of iterations, the fitness value has not been significantly improved within a certain number of generations, etc. When these conditions are met, the iteration is stopped. After the iteration is completed, the individual with the highest fitness value is output as the best hyperparameter combination to complete the automatic hyperparameter optimization.

[0077] Furthermore, the converged super-resolution model is optimized by using the optimal hyperparameter combination to obtain an optimized super-resolution model. Step S500 of the present application further includes:

[0078] Step S510: applying the optimal hyperparameter combination to the converged super-resolution model;

[0079] Step S520: collecting new training data as a new training data set to train the converged super-resolution model;

[0080] Step S530: using the new training validation set to evaluate the converged super-resolution model, calculating performance indicators to evaluate the performance of the new data, and obtaining evaluation results;

[0081] Step S540: Analyze the evaluation result to obtain the analysis result;

[0082] Step S550: Optimizing the converged super-resolution model based on the analysis result to obtain an optimized super-resolution model.

[0083] Optionally, to ensure the accuracy of the model, the converged super-resolution model needs to be retrained, evaluated, and optimized based on the new data. Specifically, the best hyperparameter combination previously obtained by the genetic algorithm is applied to the converged super-resolution model. Collect new training data, which can come from different sources or have different characteristics. Ensuring the quality and diversity of new data is crucial to improving the generalization ability of the model. Preprocess and organize the collected new data into a data set suitable for model training, that is, a new training data set, and use the new training data set to retrain the converged super-resolution model, that is, adjust the weights and parameters of the model according to the new data characteristics to adapt to the new data distribution. Prepare a new training validation set independent of the new training data set to evaluate the retrained super-resolution model, and calculate performance indicators to evaluate the performance of the new data. Record and organize the evaluation results, analyze the evaluation results, and analyze how the model performs on the new data and whether there is room for improvement. Based on the analysis results, the converged super-resolution model is optimized and the above optimization process is repeated until the performance of the model on the new data meets the expected requirements or can no longer be improved through optimization. Finally, the optimization process is stopped and the current model is saved as the optimized super-resolution model. This model will have better generalization ability and higher performance.

[0084] Through the technical solutions of the above embodiments, the LIPF-based super-resolution method for key space weather parameters provided by the present application solves the technical problems that the image super-resolution algorithm existing in the prior art has insufficient model generalization ability, high computational complexity, low algorithm practicability and efficiency when applied to the reconstruction of thermospheric atmospheric parameters, and due to the complexity and diversity of space weather data, it is difficult for the existing algorithms to effectively adapt to the needs of thermospheric atmospheric parameter reconstruction under different weather conditions, resulting in poor prediction effects. The present application provides an efficient, accurate and adaptable method for reconstructing thermospheric atmospheric parameters, which can effectively improve the resolution and accuracy of thermospheric atmospheric parameters, while reducing computational complexity and improving the practicability and efficiency of the algorithm.

[0085] Embodiment 2

[0086] Based on the same inventive concept as the LIPF-based super-resolution method for key space weather parameters in the aforementioned embodiment, Figure 2 As shown, the present application provides a space weather key parameter super-resolution system based on LIPF, the system comprising:

[0087] The parameter collection module 11 is used to determine the target area, collect key space weather parameters based on the target area, and obtain the key space weather parameters to be processed;

[0088] A data output module 12 is used to perform data processing according to the key space weather parameters to be processed to obtain the key space weather parameters, and output dual-version data respectively using an ionosphere-thermosphere numerical simulation model, wherein the dual-version data includes a high-precision data stream and a low-precision data stream;

[0089] A model building module 13 is used to build a super-resolution model based on the LIIF algorithm, and perform supervised training and cross-validation on the super-resolution model based on the dual-version data to obtain a converged super-resolution model;

[0090] The parameter optimization module 14 is used to retrieve and optimize the hyperparameters of the super-resolution model based on a genetic algorithm to obtain an optimal hyperparameter combination;

[0091] The model optimization module 15 is used to perform model optimization of the converged super-resolution model through the optimal super-parameter combination to obtain an optimized super-resolution model.

[0092] Furthermore, the data output module 12 is further configured to perform the following steps:

[0093] Performing data cleaning and data preprocessing on the key space weather parameters to be processed to obtain preprocessing key parameters;

[0094] Performing data classification and labeling on the preprocessing key parameters, performing feature extraction, labeling the preprocessing key parameters, and obtaining labeled key parameters;

[0095] Correct and eliminate the labeled key parameters to obtain space weather key parameters;

[0096] The key space weather parameters are input into the ionosphere-thermosphere numerical simulation model, and dual-version data is output.

[0097] Furthermore, the model building module 13 is also used to perform the following steps:

[0098] Select the LIIF algorithm and design the network structure of the super-resolution model based on the LIIF algorithm, including an encoder and a decoder;

[0099] Determine a loss function and select an optimizer to update the parameters of the super-resolution model;

[0100] After the parameters are updated, a super-resolution model is obtained.

[0101] Furthermore, the model building module 13 is also used to perform the following steps:

[0102] Selecting excellent samples from the dual-version data as a sample data set;

[0103] Obtaining a preset data division ratio, and dividing the sample data set into a training data set and a verification data set according to the preset data division ratio;

[0104] Performing supervised training on the super-resolution model through the training data set, and when the model output result tends to converge, verifying the output result of the super-resolution model through the verification data set;

[0105] A preset model verification accuracy index is obtained, and when the model accuracy of the super-resolution model meets the preset model verification accuracy index, a converged super-resolution model is obtained.

[0106] Furthermore, the model building module 13 is also used to perform the following steps:

[0107] Divide the training data set into multiple non-overlapping sub-training sets;

[0108] Arbitrarily select the first sub-training set as a sub-validation data set, and the remaining sub-training sets as sub-training data sets;

[0109] Cross-training the first converged super-resolution model according to the sub-validation data set and the sub-training data set to obtain a first performance indicator;

[0110] Iterate the sub-training set to obtain the performance index corresponding to the sub-training set and calculate the mean to obtain the overall performance result of the first convergence super-resolution model;

[0111] An analysis is performed based on the overall performance results to obtain a converged super-resolution model.

[0112] Furthermore, the parameter optimization module 14 is also used to perform the following steps:

[0113] Set individuals and the number of individuals, and randomly generate a set of hyperparameters for the individuals as initial solutions;

[0114] Using the super-resolution model and the training data set, training and verifying the hyper-parameter combination to obtain individual fitness;

[0115] Sorting based on the individual fitness, setting a fitness threshold, and selecting the individual whose individual fitness satisfies the fitness threshold as the parent individual;

[0116] Performing a crossover operation on the parent individuals to generate new offspring individuals, performing a mutation operation based on the new offspring individuals, combining the new offspring individuals with the parent individuals into a new hyperparameter combination, trimming the new hyperparameter combination according to the fitness threshold, and selecting the individuals with high fitness to retain;

[0117] The termination condition is preset. When the termination condition is met, the iteration is stopped and the optimal individual is output as the optimal hyperparameter combination.

[0118] Furthermore, the model optimization module 15 is also used to perform the following steps:

[0119] Based on the optimal hyperparameter combination, the converged super-resolution model is applied;

[0120] Collecting new training data as a new training data set to train the converged super-resolution model;

[0121] Using the new training validation set to evaluate the converged super-resolution model, calculating performance indicators to evaluate the performance of the new data, and obtaining evaluation results;

[0122] Analyze the evaluation result to obtain an analysis result;

[0123] The converged super-resolution model is optimized based on the analysis result to obtain an optimized super-resolution model.

[0124] Through the above detailed description of the space weather key parameter super-resolution method based on LIPF in this specification, those skilled in the art can clearly understand the space weather key parameter super-resolution method and system based on LIPF in this embodiment. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0125] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. The super-resolution method of key space weather parameters based on LIPF is characterized by: The LIPF is a method for learning continuous physical field representation using local implicit image functions. The method is based on LIIF and maps each pixel in the input image to a continuous representation space by learning a local implicit function, thereby achieving efficient representation and processing of the image. The LIPF-based space weather key parameter super-resolution method includes: Determine a target area, collect key space weather parameters based on the target area, and obtain key space weather parameters to be processed; Performing data processing according to the key space weather parameters to be processed to obtain the key space weather parameters, and using an ionosphere-thermosphere numerical simulation model to output dual-version data, respectively, wherein the dual-version data includes a high-precision data stream and a low-precision data stream; A super-resolution model is built based on the LIIF algorithm, and supervised training and cross-validation are performed on the super-resolution model based on the dual-version data to obtain a converged super-resolution model, which includes: building a network structure capable of learning mapping from low-resolution images to high-resolution images based on the LIIF algorithm, including an encoder and a decoder; wherein the encoder extracts features from the input low-resolution image, including: using a convolutional neural network as an encoder, which contains multiple convolutional layers, activation functions and pooling layers or stride convolutions to reduce the spatial size of the feature map, the input layer receives the low-resolution image as input, multiple convolutional layers extract image features, and a nonlinear activation function is applied after each convolutional layer, and the pooling layer or stride convolution reduces The spatial size of the feature map captures more global information at a higher level; the decoder reconstructs a high-resolution image from the features extracted by the encoder, wherein the decoder outputs a coordinate-related feature vector, which predicts high-resolution pixel values ​​at any coordinate, including: performing feature mapping, converting the output feature map of the encoder into a form suitable for use by the LIIF algorithm, the LIIF layer predicts high-resolution pixel values ​​from input low-resolution coordinates and features, the LIIF layer accepts low-resolution coordinates and feature mapping as input, and outputs pixel value predictions at high-resolution coordinates, wherein the LIIF layer includes a multi-layer perceptron; further, interpolating or averaging the LIIF to complete the construction of the network structure; Retrieve and optimize the hyperparameters of the super-resolution model based on genetic algorithms to obtain the best hyperparameter combination; The converged super-resolution model is optimized by using the optimal super-parameter combination to obtain an optimized super-resolution model.

2. The method for super-resolution of key space weather parameters based on LIPF as claimed in claim 1, characterized in that: Data processing is performed according to the key space weather parameters to be processed to obtain the key space weather parameters, and dual-version data is outputted respectively using an ionosphere-thermosphere numerical simulation model, wherein the dual-version data includes a high-precision data stream and a low-precision data stream, and the method further includes: Performing data cleaning and data preprocessing on the key space weather parameters to be processed to obtain preprocessing key parameters; Performing data classification and labeling on the preprocessing key parameters, performing feature extraction, labeling the preprocessing key parameters, and obtaining labeled key parameters; Correct and eliminate the labeled key parameters to obtain space weather key parameters; The key space weather parameters are input into the ionosphere-thermosphere numerical simulation model, and dual-version data is output.

3. The method for super-resolution of key space weather parameters based on LIPF as claimed in claim 1, characterized in that: A super-resolution model is constructed based on the LIIF algorithm, and supervised training and cross-validation are performed on the super-resolution model based on the dual-version data to obtain a converged super-resolution model. The method further includes: Determine a loss function and select an optimizer to update the parameters of the super-resolution model; After the parameters are updated, a super-resolution model is obtained.

4. The method for super-resolution of key space weather parameters based on LIPF as claimed in claim 1, characterized in that: Based on the dual-version data, supervised training and cross-validation are performed on the super-resolution model to obtain a converged super-resolution model. The method further includes: Selecting excellent samples from the dual-version data as a sample data set; Obtaining a preset data division ratio, and dividing the sample data set into a training data set and a verification data set according to the preset data division ratio; Performing supervised training on the super-resolution model through the training data set, and when the model output result tends to converge, verifying the output result of the super-resolution model through the verification data set; A preset model verification accuracy index is obtained, and when the model accuracy of the super-resolution model meets the preset model verification accuracy index, a converged super-resolution model is obtained.

5. The method for super-resolution of key space weather parameters based on LIPF as claimed in claim 4, characterized in that: The method further comprises: Divide the training data set into multiple non-overlapping sub-training sets; Arbitrarily select the first sub-training set as a sub-validation data set, and the remaining sub-training sets as sub-training data sets; Cross-training the first converged super-resolution model according to the sub-validation data set and the sub-training data set to obtain a first performance indicator; Iterate the sub-training set to obtain the performance index corresponding to the sub-training set and calculate the mean to obtain the overall performance result of the first convergence super-resolution model; An analysis is performed based on the overall performance results to obtain a converged super-resolution model.

6. The method for super-resolution of key space weather parameters based on LIPF as claimed in claim 5, characterized in that: Retrieving and optimizing the hyperparameters of the super-resolution model based on a genetic algorithm to obtain an optimal hyperparameter combination, the method further comprising: Set individuals and the number of individuals, and randomly generate a set of hyperparameters for the individuals as initial solutions; Using the super-resolution model and the training data set, training and verifying the hyper-parameter combination to obtain individual fitness; Sorting based on the individual fitness, setting a fitness threshold, and selecting the individual whose individual fitness satisfies the fitness threshold as the parent individual; Performing a crossover operation on the parent individuals to generate new offspring individuals, performing a mutation operation based on the new offspring individuals, combining the new offspring individuals with the parent individuals into a new hyperparameter combination, trimming the new hyperparameter combination according to the fitness threshold, and selecting the individuals with high fitness to retain; The termination condition is preset. When the termination condition is met, the iteration is stopped and the optimal individual is output as the optimal hyperparameter combination.

7. The method for super-resolution of key space weather parameters based on LIPF as claimed in claim 1, characterized in that: The converged super-resolution model is optimized by performing the model optimization of the converged super-resolution model through the optimal super-parameter combination to obtain an optimized super-resolution model. The method further includes: Based on the optimal hyperparameter combination, the converged super-resolution model is applied; Collecting new training data as a new training data set to train the converged super-resolution model; Using the new training validation set to evaluate the converged super-resolution model, calculating performance indicators to evaluate the performance of the new data, and obtaining evaluation results; Analyze the evaluation result to obtain an analysis result; The converged super-resolution model is optimized based on the analysis result to obtain an optimized super-resolution model.

8. The space weather key parameter super-resolution system based on LIPF is characterized by: The system is used to implement the LIPF-based space weather key parameter super-resolution method according to any one of claims 1 to 7, comprising: A parameter collection module is used to determine a target area, collect key space weather parameters based on the target area, and obtain key space weather parameters to be processed; A data output module, configured to perform data processing according to the key space weather parameters to be processed to obtain the key space weather parameters, and output dual-version data using an ionosphere-thermosphere numerical simulation model, wherein the dual-version data includes a high-precision data stream and a low-precision data stream; A model building module, used to build a super-resolution model based on the LIIF algorithm, and perform supervised training and cross-validation on the super-resolution model based on the dual-version data to obtain a converged super-resolution model; The parameter optimization module is used to retrieve and optimize the hyperparameters of the super-resolution model based on the genetic algorithm to obtain the best hyperparameter combination; The model optimization module is used to perform model optimization of the converged super-resolution model through the optimal super-parameter combination to obtain an optimized super-resolution model.

Images