Wind cloud stationary satellite cloud picture spatial resolution improving method based on Real-ESRGAN network

Abstract

The invention discloses a wind cloud stationary satellite cloud picture spatial resolution improving method based on a Real-ESRGAN network, and the method comprises the steps: obtaining FY4B cloud picture data, and carrying out the preprocessing of the FY4B cloud picture data; performing fuzzy judgment on the FY4B cloud picture data to obtain a to-be-corrected cloud picture, and constructing a wind cloud stationary satellite cloud picture spatial resolution improvement model composed of a generator and a discriminator; and optimizing the wind cloud geostationary satellite cloud picture spatial resolution improving model through minimizing a loss function, inputting the to-be-corrected cloud picture into the optimized wind cloud geostationary satellite cloud picture spatial resolution improving model, and outputting a high-resolution cloud picture. The method not only can improve the precision of improving the spatial resolution of the cloud picture of the wind cloud geostationary satellite, but also has better interpretability, and can be directly applied to a system for improving the spatial resolution of the cloud picture of the wind cloud geostationary satellite.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of meteorological cloud maps, and in particular to a wind cloud stationary satellite cloud map spatial resolution enhancement method based on a Real-ESRGAN network. BACKGROUND

[0002] With the development of meteorological undertakings, the demand for high-precision high-resolution data is increasing, whether it is for improving the model or for meteorological data analysis and mining, the existing satellite data high-resolution channels are often visible light channels, the wind cloud 4B (FY4B) satellite fast scanning channel can reach 250m resolution, however, the infrared channel resolution is low, especially the 10.8um channel commonly used for various cloud microphysical quantity inversion has only 4km resolution, in weather modification, the precision required by the business is often high, and the operation point and the affected area are often a local range, therefore, a method is needed to enhance the resolution of FY4B stationary satellite cloud maps and their products, and from the hardware, the resolution is enhanced, which not only means that it is difficult to achieve technically, but also means that the cost is increased, so it is very meaningful to carry out super-resolution reconstruction of satellite data.

[0003] Currently, some scholars have carried out super-resolution reconstruction (downscaling) of meteorological data, but most of them are super-resolution reconstruction of reanalysis data, hourly precipitation data and other ground / model products, and there are few existing studies on super-resolution reconstruction of meteorological satellite data, especially super-resolution reconstruction of stationary satellite data in China. The existing super-resolution reconstruction of meteorological data often uses two methods, one is traditional interpolation, which can generate high-resolution data efficiently, but cannot well fit the real distribution of the data; the other is a deep learning method, and in recent years, with the development of deep learning, great achievements have been made in the field of super-resolution reconstruction, and many scholars have also adopted deep learning-based super-resolution reconstruction, which is better than the traditional interpolation method, but still has problems such as low image quality and insufficient detail restoration. Therefore, it is considered to use the Real-ESRGAN network to realize the super-resolution reconstruction of the wind cloud 4B-10.8um channel. SUMMARY

[0004] The purpose of the present application is to provide a wind cloud stationary satellite cloud map spatial resolution enhancement method based on a Real-ESRGAN network.

[0005] To achieve the above purpose, the present application is implemented according to the following technical solutions: The present application comprises the following steps: obtaining FY4B cloud map data, and preprocessing the FY4B cloud map data; The FY4B cloud image data is subjected to fuzzy judgment to obtain a cloud image to be corrected, and a Fengyun geostationary satellite cloud image spatial resolution improvement model composed of a generator and a discriminator is constructed. The Fengyun geostationary satellite cloud image spatial resolution improvement model is optimized by minimizing a loss function, and the cloud image to be corrected is input into the optimized Fengyun geostationary satellite cloud image spatial resolution improvement model to output a high-resolution cloud image.

[0006] Further, the method for fuzzy judgment on the FY4B cloud image data comprises the following steps. The FY4B cloud image set is taken as a decision object, and a set of field experts is taken as decision makers according to the FY4B cloud image data, and each expert independently gives a fuzzy judgment matrix, i.e., an expert fuzzy matrix. A lower bound matrix and an upper bound matrix of group judgment are calculated according to the expert fuzzy judgment matrix, and an additive consistent fuzzy judgment matrix is solved under a constraint condition to obtain a consistent matrix, wherein the constraint condition is that the distance between the fuzzy judgment matrix and the lower bound matrix is minimum, and the fuzzy judgment matrix is greater than the lower bound matrix and less than the upper bound matrix. The consistency distance of the expert fuzzy matrix and the consistent matrix is calculated, and if the consistency distance is less than a consistency threshold value, the consistent matrix of all experts after correction is aggregated into a collective fuzzy judgment matrix through weighted arithmetic averaging. Otherwise, the absolute difference value of each non-diagonal element of the expert fuzzy matrix and the corresponding element of the consistent matrix is calculated, the element position with an absolute deviation greater than a deviation threshold value is identified, and a to-be-corrected element set is formed according to the element position. The to-be-corrected element of the current iteration is corrected into a non-diagonal correction element of the next iteration, so that the non-diagonal correction element approaches the corresponding element of the consistent matrix, and the complementary elements of the to-be-corrected element are simultaneously corrected according to the complementarity of the fuzzy judgment matrix, and the elements outside the to-be-corrected element set remain unchanged. The lower bound matrix and the upper bound matrix are updated by using the expert fuzzy matrix of the corrected non-diagonal correction element, the iteration number is increased by one, and the additive consistent fuzzy judgment matrix is solved again, and the iteration is continuously performed until all the to-be-corrected element set is traversed to obtain a consistent judgment matrix. The multiple consistent judgment matrices are aggregated into a collective fuzzy judgment matrix through weighted arithmetic averaging, the sorting weight vector of the FY4B cloud image is calculated, and the FY4B cloud image with a sorting weight vector greater than a sorting threshold vector is taken as the cloud image to be corrected.

[0007] Further, the method of the generator comprises the following steps. The generator is composed of a 3-layer residual feature distillation module, the residual feature distillation module combines a residual dense network multi-branch structure, introduces a hollow convolution layer, reduces the parameter amount of the model, and increases the feature representation of the lightweight network, and the expression of the residual feature distillation module is: wherein is input data, is a first layer convolutional network, the convolution kernel is 3*3 in size, the input channel is 1, the output channel is 64, the step is 1, and the padding is 1; is a first layer residual feature distillation module, is a feature extracted by the first layer convolutional network; is a second layer residual feature distillation module, is a feature extracted by the second layer residual feature distillation module; is a third layer residual feature distillation module, is a feature extracted by the third layer residual feature distillation module; is a second layer convolutional network, the convolution kernel is 3*3 in size, the input channel is 64, the output channel is 64, the step is 1, and the padding is 1, is a feature extracted by the second layer convolutional network; is an up-sampling module, is a feature obtained by the up-sampling network; is a third layer convolutional network, the convolution kernel is 3*3 in size, the input channel is 64, the output channel is 1, the step is 1, and the padding is 1; is a high-resolution output obtained by the third layer convolutional network; is a feature extracted by the first layer convolutional network; The up-sampling module is used to process the high-resolution output, and the expression is as follows: wherein is input feature, represents a convolutional layer, the convolution kernel is 3*3, represents a feature obtained by the convolutional layer; is an activation function, is a feature obtained by the activation function; is a pixel reshuffling function, is an up-sampling ratio, The features are obtained after the pixel reshuffling function. To simplify the expression, the generator is defined as G. Then, to update the high-resolution output to Fout, we have the following expression: in This is the updated high-resolution output.

[0008] Furthermore, the method of the discriminator includes: The discriminator uses a UNET network structure. The high-resolution output from the generator is fed into the discriminator to obtain the probability. The result obtained after inputting real data into the discriminator is expressed as follows: in For the discriminator model, For the high-resolution output of the generator. The result after real data is fed into the discriminator. for The probability after being placed in the discriminator. This is the residual characteristic distillation module.

[0009] Furthermore, the method for optimizing the spatial resolution improvement model of the Fengyun geostationary satellite cloud image by minimizing the loss function includes: The loss function chosen is the adversarial generative loss function, which includes a generator loss function and a discriminator loss function, expressed as follows: in Let be the generator loss function. Let the discriminator loss function be... The loss function for the discriminator to identify true values. The loss function for the discriminator to identify something as false. For pixel loss function, The crossover loss function; An adaptive learning rate strategy is adopted to set the initial learning rate, momentum parameter, and decay weight; The gradient of the loss function with respect to the network parameters is calculated using the backpropagation algorithm, and the update formula is as follows: in Let be the network parameters for the t-th iteration. For the network parameters in the (t+1)th iteration, For learning rate, This is the first-order moment estimate after bias correction in the t-th iteration. This is the first and second moment estimate after bias correction for the t-th iteration. Let be the error constant; Continuously reduce the loss function value until the optimal state is reached.

[0010] The beneficial effects of this invention are: This invention is a method for enhancing the spatial resolution of geostationary satellite cloud images based on Real-ESRGAN networks. Compared with existing technologies, this invention has the following technical advantages: This invention improves the accuracy of user intent recognition in multi-dimensional human-computer interaction scenarios through preprocessing, fuzzy judgment, generator and discriminator construction, model building, and model optimization steps. By introducing a residual feature distillation module and a fuzzy judgment mechanism with local correction, it significantly enhances the ability to restore cloud image details while reducing the number of parameters. The multi-component loss function optimization strategy effectively improves the accuracy and visual quality of super-resolution reconstruction, making it more suitable for the refined processing needs of infrared channel cloud images in meteorological operations. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the steps of the method for improving the spatial resolution of geostationary satellite cloud images based on Real-ESRGAN networks according to the present invention. Figure 2 The high-resolution cloud image shown in this embodiment of the invention is shown in Figure 1, where a is the collected FY4B cloud image data and b is the final output high-resolution FY4B satellite cloud image. Detailed Implementation

[0012] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0013] The spatial resolution enhancement method for geostationary satellite cloud images based on Real-ESRGAN networks of this invention includes the following steps: like Figure 1 As shown, this embodiment includes the following steps: Acquire FY4B cloud map data and preprocess the FY4B cloud map data; In the actual evaluation, FY4B satellite imagery data was obtained from a certain network resource pool. The FY4B data has a spatial resolution of 4km and a temporal resolution of 15 minutes. FY4B satellite imagery data for the entire year of 2022-2023 was collected. After filtering and removing low-quality data, 2240 satellite data points were obtained, which were used to construct the training dataset required for model training. The test set randomly selected data from 2024, approximately 20% (559 data points) of the training dataset. Python was used to crop the obtained datasets (training and test sets), with a cropping matrix size of 256*256. First, lower-resolution data was created by downsampling the dataset to 64*64, resulting in a low-resolution dataset (64*64) and an original resolution dataset (256*256). Finally, the data was saved in NPZ format, such as... Figure 2 As shown in 'a'; Preprocessing: Satellite image data values ​​range from 0 to 6535, resulting in a relatively wide sample distribution. However, by examining the distribution range of each sample in the dataset, it was found that the values ​​are mainly distributed between 1900 and 4000. Therefore, values ​​less than 1900 were set to 1900, and values ​​greater than 4000 were set to 4000. Standardizing the value range of the satellite images before normalization ensures a reasonable data distribution, facilitating model training iterations. To increase the robustness of the model, data augmentation methods were employed, including random cropping and flipping of the data. Fuzzy judgment is performed on the FY4B cloud image data to obtain the cloud image to be corrected, and a spatial resolution improvement model for Fengyun geostationary satellite cloud image composed of a generator and a discriminator is constructed. In the actual evaluation, the upsampling ratio was set to 4. Python was used to call the generator model, load the parameters trained by the generator model, and use the generator model to perform super-resolution reconstruction on the standard npz format input data. After reconstruction, a new super-resolution data npz_SR data was obtained. The npz_SR data is denormalized using an inverse normalization function, and finally a high-resolution FY4B satellite cloud image is output, such as... Figure 2 As shown in b; The spatial resolution enhancement model of the Fengyun geostationary satellite cloud image is optimized by minimizing the loss function. The cloud image to be corrected is input into the optimized spatial resolution enhancement model of the Fengyun geostationary satellite cloud image, and a high-resolution cloud image is output.

[0014] In this embodiment, the method for performing fuzzy judgment on the FY4B cloud map data includes: Based on the FY4B cloud map data, the FY4B cloud map set is used as the decision object, and the domain expert set is used as the decision-maker. Each expert independently provides a fuzzy judgment matrix, which is the expert fuzzy matrix. Calculate the lower and upper bound matrices of the group judgment based on the fuzzy judgment matrix of the experts, and solve for an additively consistent fuzzy judgment matrix under the constraints to obtain the consistency matrix; where the constraints are that the distance between the fuzzy judgment matrix and the lower bound matrix is ​​minimized, and the fuzzy judgment matrix is ​​greater than the lower bound matrix and less than the upper bound matrix. Calculate the consistency distance between the expert fuzzy matrix and the consistency matrix. If the consistency distance is less than the consistency threshold, then integrate all the expert-corrected consistency matrices into a collective fuzzy judgment matrix by weighted arithmetic average. Conversely, calculate the absolute difference between each off-diagonal element of the expert fuzzy matrix and the corresponding element of the consistency matrix, identify the position of the element whose absolute deviation is greater than the deviation threshold, and construct a set of elements to be corrected based on the element positions. The elements to be corrected in the current iteration are corrected to the off-diagonal correction elements in the next iteration, so that the off-diagonal correction elements are infinitely close to the corresponding elements of the consistency matrix. At the same time, based on the complementarity of the fuzzy judgment matrix, the complementary elements of the elements to be corrected are corrected simultaneously, while the elements outside the set of elements to be corrected remain unchanged. Update the upper and lower bound matrices using the expert fuzzy matrix that corrects off-diagonal elements, increment the iteration count by one, resolve the additively consistent fuzzy judgment matrix, and iterate until all sets of elements to be corrected are traversed to obtain the consistent judgment matrix. Multiple consistency judgment matrices are integrated into a collective fuzzy judgment matrix by weighted arithmetic average. The ranking weight vector of the FY4B cloud map is calculated, and the FY4B cloud map with the ranking weight vector greater than the ranking threshold vector is taken as the cloud map to be corrected.

[0015] In this embodiment, the generator method includes: A generator is constructed using a three-layer residual feature distillation module. This module combines a dense residual network with a multi-branch structure and introduces dilated convolutional layers to reduce the number of model parameters while increasing the feature representation of a lightweight network. The expression for the residual feature distillation module is as follows: in For input data, This is the first layer of the convolutional network, with a kernel size of 3*3, 1 input channel, 64 output channels, a stride of 1, and padding of 1. This is the first layer of residual characteristic distillation module. These are the features extracted after passing through the first convolutional network layer; This is the second-layer residual characteristic distillation module. These are the features extracted by the second-layer residual feature distillation module; This is the third-layer residual characteristic distillation module. These are the features extracted by the third-layer residual feature distillation module; This is the second convolutional layer, with a 3x3 kernel size, 64 input channels, 64 output channels, a stride of 1, and padding of 1. These are the features extracted after passing through the second convolutional network. For upsampling module, These are the features obtained after upsampling the network; This is the third layer of the convolutional network, with a kernel size of 3*3, 64 input channels, 1 output channel, a stride of 1, and padding of 1. This is the high-resolution output obtained after passing through the third convolutional network layer; These are the features extracted through the first layer of the convolutional network; The high-resolution output is processed using an upsampling module, as shown in the following expression: in For the input features, This represents a convolutional layer with a 3x3 kernel. This represents the features obtained after passing through the convolutional layer; For activation function, These are the features obtained after passing through the activation function; For pixel reshuffling function, This is the upsampling ratio. The features are obtained after the pixel reshuffling function. To simplify the expression, the generator is defined as G. Then, to update the high-resolution output to Fout, we have the following expression: in This is the updated high-resolution output.

[0016] In this embodiment, the method of the discriminator includes: The discriminator uses a UNET network structure. The high-resolution output from the generator is fed into the discriminator to obtain the probability. The result obtained after inputting real data into the discriminator is expressed as follows: in For the discriminator model, For the high-resolution output of the generator. The result after real data is fed into the discriminator. for The probability after being placed in the discriminator. This is the residual characteristic distillation module.

[0017] In this embodiment, the method for optimizing the spatial resolution enhancement model of the Fengyun geostationary satellite cloud image by minimizing the loss function includes: The loss function chosen is the adversarial generative loss function, which includes a generator loss function and a discriminator loss function, expressed as follows: in Let be the generator loss function. Let the discriminator loss function be... The loss function for the discriminator to identify true values. The loss function for the discriminator to identify something as false. For pixel loss function, The crossover loss function; An adaptive learning rate strategy is adopted to set the initial learning rate, momentum parameter, and decay weight; The gradient of the loss function with respect to the network parameters is calculated using the backpropagation algorithm, and the update formula is as follows: in Let be the network parameters for the t-th iteration. For the network parameters in the (t+1)th iteration, For learning rate, This is the first-order moment estimate after bias correction in the t-th iteration. This is the first and second moment estimate after bias correction for the t-th iteration. Let be the error constant; Continuously reduce the loss function value until the optimal state is reached.

[0018] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for improving the spatial resolution of geostationary satellite cloud images based on Real-ESRGAN networks, characterized in that, Includes the following steps: Acquire FY4B cloud map data and preprocess the FY4B cloud map data; Fuzzy judgment is performed on the FY4B cloud image data to obtain the cloud image to be corrected, and a spatial resolution improvement model for Fengyun geostationary satellite cloud image composed of a generator and a discriminator is constructed. The spatial resolution enhancement model of the Fengyun geostationary satellite cloud image is optimized by minimizing the loss function. The cloud image to be corrected is input into the optimized spatial resolution enhancement model of the Fengyun geostationary satellite cloud image, and a high-resolution cloud image is output.

2. The method for improving the spatial resolution of geostationary satellite cloud images based on Real-ESRGAN networks according to claim 1, characterized in that, The method for performing fuzzy judgment on the FY4B cloud map data includes: Based on the FY4B cloud map data, the FY4B cloud map set is used as the decision object, and the domain expert set is used as the decision-maker. Each expert independently provides a fuzzy judgment matrix, which is the expert fuzzy matrix. Calculate the lower and upper bound matrices of the group judgment based on the fuzzy judgment matrix of the experts, and solve for an additively consistent fuzzy judgment matrix under the constraints to obtain the consistency matrix; where the constraints are that the distance between the fuzzy judgment matrix and the lower bound matrix is ​​minimized, and the fuzzy judgment matrix is ​​greater than the lower bound matrix and less than the upper bound matrix. Calculate the consistency distance between the expert fuzzy matrix and the consistency matrix. If the consistency distance is less than the consistency threshold, then integrate all the expert-corrected consistency matrices into a collective fuzzy judgment matrix by weighted arithmetic average. Conversely, calculate the absolute difference between each off-diagonal element of the expert fuzzy matrix and the corresponding element of the consistency matrix, identify the position of the element whose absolute deviation is greater than the deviation threshold, and construct a set of elements to be corrected based on the element positions. The elements to be corrected in the current iteration are corrected to the off-diagonal correction elements in the next iteration, so that the off-diagonal correction elements are infinitely close to the corresponding elements of the consistency matrix. At the same time, based on the complementarity of the fuzzy judgment matrix, the complementary elements of the elements to be corrected are corrected simultaneously, while the elements outside the set of elements to be corrected remain unchanged. Update the upper and lower bound matrices using the expert fuzzy matrix that corrects off-diagonal elements, increment the iteration count by one, resolve the additively consistent fuzzy judgment matrix, and iterate until all sets of elements to be corrected are traversed to obtain the consistent judgment matrix. Multiple consistency judgment matrices are integrated into a collective fuzzy judgment matrix by weighted arithmetic average. The ranking weight vector of the FY4B cloud map is calculated, and the FY4B cloud map with the ranking weight vector greater than the ranking threshold vector is taken as the cloud map to be corrected.

3. The method for improving the spatial resolution of geostationary satellite cloud images based on Real-ESRGAN networks according to claim 1, characterized in that, The generator method includes: A generator is constructed using a three-layer residual feature distillation module. This module combines a dense residual network with a multi-branch structure and introduces dilated convolutional layers to reduce the number of model parameters while increasing the feature representation of a lightweight network. The expression for the residual feature distillation module is as follows: in For input data, This is the first layer of the convolutional network, with a kernel size of 3*3, 1 input channel, 64 output channels, a stride of 1, and padding of 1. This is the first layer of residual characteristic distillation module. These are the features extracted after passing through the first convolutional network layer; This is the second-layer residual characteristic distillation module. These are the features extracted by the second-layer residual feature distillation module; This is the third-layer residual characteristic distillation module. These are the features extracted by the third-layer residual feature distillation module; This is the second layer of the convolutional network, with a kernel size of 3x3, 64 input channels, 64 output channels, a stride of 1, and padding of 1. These are the features extracted after passing through the second convolutional network. For upsampling module, These are the features obtained after upsampling the network; This is the third layer of the convolutional network, with a kernel size of 3*3, 64 input channels, 1 output channel, a stride of 1, and padding of 1. This is the high-resolution output obtained after passing through the third convolutional network layer; These are the features extracted through the first layer of the convolutional network; The high-resolution output is processed using an upsampling module, as shown in the following expression: in For the input features, This represents a convolutional layer with a 3x3 kernel. This represents the features obtained after passing through the convolutional layer; For activation function, These are the features obtained after passing through the activation function; For pixel reshuffling function, This is the upsampling ratio. The features are obtained after the pixel reshuffling function. To simplify the expression, the generator is defined as G. Then, to update the high-resolution output to Fout, we have the following expression: in This is the updated high-resolution output.

4. The method for improving the spatial resolution of geostationary satellite cloud images based on Real-ESRGAN networks according to claim 1, characterized in that, The method of the discriminator includes: The discriminator uses a UNET network structure. The high-resolution output from the generator is fed into the discriminator to obtain the probability. The result obtained after inputting real data into the discriminator is expressed as follows: in For the discriminator model, For the high-resolution output of the generator. The result after real data is fed into the discriminator. for The probability after being placed in the discriminator. This is the residual characteristic distillation module.

5. The method for improving the spatial resolution of geostationary satellite cloud images based on Real-ESRGAN networks according to claim 1, characterized in that, The method for optimizing the spatial resolution enhancement model of the Fengyun geostationary satellite cloud image by minimizing the loss function includes: The loss function chosen is the adversarial generative loss function, which includes a generator loss function and a discriminator loss function, expressed as follows: in Let be the generator loss function. Let the discriminator loss function be... The loss function for the discriminator to identify true values. The loss function for the discriminator to identify something as false. For pixel loss function, The crossover loss function; An adaptive learning rate strategy is adopted to set the initial learning rate, momentum parameter, and decay weight; The gradient of the loss function with respect to the network parameters is calculated using the backpropagation algorithm, and the update formula is as follows: in Let be the network parameters for the t-th iteration. For the network parameters in the (t+1)th iteration, For learning rate, This is the first-order moment estimate after bias correction in the t-th iteration. This is the first and second moment estimate after bias correction for the t-th iteration. Let be the error constant; Continuously reduce the loss function value until the optimal state is reached.

Images