Meteorological satellite cloud picture super-division processing method and system based on pixel convolutional network

Abstract

The invention discloses a meteorological satellite cloud picture super-division processing method and system based on a pixel convolutional network, and relates to the technical field of cloud picture super-division processing, and the method comprises the following steps: S1, obtaining a meteorological satellite cloud picture in real time; s2, preprocessing a meteorological satellite cloud picture; s3, constructing a meteorological satellite cloud picture super-division processing model based on the pixel convolutional network; and S4, carrying out real-time super-resolution processing on the meteorological satellite cloud picture. According to the method, the low-resolution FY4B meteorological satellite cloud picture is acquired and preprocessed in real time, the satellite cloud picture is analyzed into pixel data, super-division processing is conveniently carried out on a pixel-level layer, a meteorological satellite cloud picture super-division processing model based on the pixel convolutional network is researched and developed, the calculation efficiency is greatly improved, and the calculation cost is reduced. Parameters are continuously adjusted through a loss function to train a meteorological satellite cloud picture super-division processing model, the super-division effect is improved, and the problems that the meteorological satellite cloud picture processing speed is insufficient, and the fine degree is insufficient are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of cloud image super-resolution processing, and in particular to a meteorological satellite cloud image super-resolution processing method and system based on a pixel convolutional network. Background Art

[0002] A Chinese patent with application number CN202110251986.X discloses an image super-resolution processing method, an image super-resolution processing device and a storage medium, wherein the image super-resolution processing method is applied to a terminal, including: determining an image to be processed; determining an image super-resolution model based on the power operation status of the terminal and the image resolution of the image to be processed, and performing image super-resolution processing on the image to be processed based on the image super-resolution model.

[0003] Existing technologies can better balance the super-resolution processing effect and the power consumption of the terminal system during image super-resolution processing, bringing users a good user experience. However, when applied to meteorological satellite cloud maps, the cloud maps are not efficient enough during extraction and processing, and cannot respond quickly to new cloud map data. If you want to accurately predict meteorological changes, especially for refined forecasts at the township level, you need to improve the resolution of the cloud maps and the frequency of updating the cloud maps to achieve good super-resolution results. This process is difficult to maintain excellent precision and high clarity, so refined early warning forecasts become difficult. Summary of the invention

[0004] The purpose of the present invention is to provide a meteorological satellite cloud image super-resolution processing method and system based on a pixel convolutional network to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a method for super-resolution processing of meteorological satellite cloud images based on pixel convolutional network, comprising the following steps: S1. Obtain low-resolution satellite cloud images from the meteorological big data cloud platform, and update the satellite cloud images every 15 minutes. After collection, process the low-resolution satellite cloud images; S2. Use the grid computing method based on the CGMS specification to locate and register the meteorological satellite cloud image of the designated area, convert it into Mercator projection, and finally parse the satellite cloud image into pixel data to facilitate super-resolution processing at the pixel level; S3. Build a meteorological satellite cloud image super-resolution processing model based on pixel convolutional network, select meteorological satellite cloud images covering various weather characteristics such as daytime, nighttime, sunny days and rainy days, compress them as training data sets, and select 10% of the uncompressed original meteorological satellite cloud images as verification data sets. Use the above data sets to perform super-resolution training on the meteorological satellite cloud image super-resolution processing model, so that the refinement and clarity of the meteorological satellite cloud images are significantly improved; S4. Deploy the meteorological satellite cloud image super-resolution processing model on the meteorological big data cloud platform, perform feature extraction on the input low-resolution meteorological satellite cloud image in real time, and directly output high-resolution satellite cloud image through pixel convolution calculation.

[0006] Preferably, the grid calculation method formula based on the CGMS specification is: In the formula, It means that the meteorological satellite points to the center of the earth where the target object is a known quantity. represents the meteorological satellite spin vector, express and The angle between the two can be measured by observing the row number of the center of the earth in the cloud map. Represents the cosine function.

[0007] Preferably, in step S3, the construction of the pixel convolutional network includes the following steps: S301, the original low-resolution meteorological satellite cloud image is subjected to three convolution layers of different sizes, and the sizes of the convolution kernels used are , , ,The feature mapping area extracted from the satellite cloud images output by each layer has different sizes, and the features of different sizes are merged by splicing to extract distortion features and eliminate compression artifacts and noise; S302, construct three layers of convolution to extract the main features of the meteorological satellite cloud map. In the first layer of convolution, the number of filters is 64 and the convolution kernel size is In the second convolution layer, the number of filters is 32 and the convolution kernel size is In the third convolution layer, the number of filters is 16 and the convolution kernel size is ; S303, the merged features and the extracted main features are fused, and then a residual dense module is used to extract rich local features as output features. The residual dense module is composed of three residual dense blocks.

[0008] Preferably, in step S303, the residual dense module is represented as: In the formula, represents the input features, represents the output features, represents the residual dense block, represents the residual scaling factor, represents the local features extracted by the first residual dense block, represents the local features extracted by the second residual dense block, Represents the local features extracted by the third residual dense block.

[0009] Preferably, the construction of the pixel convolutional network also includes the following steps: S304, the activation functions of the first and second convolution layers use the ReLU function, and the activation function of the third convolution layer uses the TanH function, and the size of the input meteorological satellite cloud image and the number of feature channels are consistent. The feature map of is the zoom factor; S305: Add the feature map The feature channels are rearranged into This area corresponds to a high-resolution meteorological satellite cloud map This operation converts the original width to Gao Wei And the number of channels is The feature map is directly converted into a width Gao Wei And the number of channels is feature map.

[0010] Preferably, the meteorological satellite cloud image super-resolution processing model is: In the formula, represents the operation based on pixel convolutional network, and Respectively represent The parameters of the layer, Indicates The convolution operation of the layer, and They represent low-resolution meteorological satellite cloud images and high-resolution meteorological satellite cloud images respectively. Represents a convolutional layer.

[0011] Preferably, in step S3, a loss function is used to continuously adjust parameters to train the meteorological satellite super-resolution processing model, and the loss function is expressed as: In the formula, represents the loss function, Represents the convolution kernel and bias parameters to be learned. represents a neural network that generates fractional pixels, and They represent low-resolution meteorological satellite cloud images and high-resolution meteorological satellite cloud images respectively. Indicates the number of parameter adjustments. , Indicates taking the absolute value.

[0012] A meteorological satellite cloud image super-resolution processing system based on a pixel convolutional network, comprising a cloud image acquisition unit, a cloud image processing unit, a model building unit and a super-resolution processing unit; The cloud image acquisition unit is used to collect and process low-resolution meteorological satellite cloud images, obtain low-resolution satellite cloud images from the meteorological big data cloud platform, and update the satellite cloud images every 15 minutes, and process the low-resolution satellite cloud images after collection; The cloud image processing unit parses the satellite cloud image into pixel data, so as to facilitate super-resolution processing at the pixel level; The model building unit builds a meteorological satellite cloud image super-resolution processing model based on a pixel convolutional network, selects meteorological satellite cloud images covering multiple weather characteristics of daytime, nighttime, sunny days, and rainy days, compresses them as training data sets, and selects 10% from the uncompressed original meteorological satellite cloud images as a verification data set, and uses the above data sets to perform super-resolution training on the meteorological satellite cloud image super-resolution processing model, so that the refinement and clarity of the meteorological satellite cloud images are significantly improved; The super-resolution processing unit performs feature extraction on the input low-resolution meteorological satellite cloud image in real time, and directly outputs the high-resolution satellite cloud image through pixel convolution calculation.

[0013] Compared with the prior art, the present invention has the following beneficial effects: The present invention acquires and pre-processes low-resolution FY4B meteorological satellite cloud images in real time, parses meteorological satellite cloud images into pixel data, facilitates super-resolution processing at the pixel level, constructs a meteorological satellite cloud image super-resolution processing model based on a pixel convolutional network, processes a meteorological satellite cloud image in less than 1 second, greatly improves the calculation efficiency, quickly responds to new meteorological satellite cloud image data, uses a loss function to continuously adjust parameters to train the meteorological satellite cloud image super-resolution processing model, improves the super-resolution effect, and the optimized super-resolution meteorological satellite cloud image has a maximum PSNR of 45.56 dB. After the model is trained, the original FY4B satellite cloud image is super-resolved, which significantly improves the degree of refinement and clarity, and solves the problems of insufficient processing speed and insufficient degree of refinement of meteorological satellite cloud images. The present invention trains a meteorological satellite cloud image super-resolution processing model by selecting a data set of meteorological satellite cloud images covering various weather characteristics of daytime, night, sunny days and rainy days, so that the model training accuracy is improved. The meteorological satellite cloud image super-resolution processing model performs 4-fold super-resolution processing on the FY4B satellite cloud image, thereby improving the super-resolution processing speed. The meteorological satellite cloud image after super-resolution processing is clearer than the meteorological satellite cloud image directly magnified by a multiple, and the super-resolution processing effect is better. The meteorological satellite cloud image after super-resolution serves for short-term meteorological disaster warning, facilitates the early identification of initial severe convective weather, and provides strong support for the prevention and control of urban waterlogging, small and medium-sized river floods, and mountain torrents and geological disasters. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A schematic diagram of the overall method flow provided by an embodiment of the present invention; Figure 2 A schematic diagram of the overall system structure provided by an embodiment of the present invention; Figure 3 A schematic diagram showing a comparison of the effects of super-resolution processing before and after provided in an embodiment of the present invention, wherein a is a local cloud map before super-resolution processing, and b is a local cloud map after super-resolution processing. DETAILED DESCRIPTION

[0015] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0016] See also Figures 1 to 3 The present invention provides a technical solution: a super-resolution processing method of meteorological satellite cloud images based on a pixel convolutional network, comprising the following steps: S1. Obtain low-resolution satellite cloud images from the meteorological big data cloud platform, and update the satellite cloud images every 15 minutes. After collection, process the low-resolution satellite cloud images; S2. Use the grid computing method based on the CGMS specification to locate and register the meteorological satellite cloud image of the designated area, convert it into Mercator projection, and finally parse the satellite cloud image into pixel data to facilitate super-resolution processing at the pixel level; S3. Build a meteorological satellite cloud image super-resolution processing model based on pixel convolutional network, select meteorological satellite cloud images covering various weather characteristics such as daytime, nighttime, sunny days and rainy days, compress them as training data sets, and select 10% of the uncompressed original meteorological satellite cloud images as verification data sets. Use the above data sets to perform super-resolution training on the meteorological satellite cloud image super-resolution processing model, so that the refinement and clarity of the meteorological satellite cloud images are significantly improved; S4. Deploy the meteorological satellite cloud image super-resolution processing model on the meteorological big data cloud platform, perform feature extraction on the input low-resolution meteorological satellite cloud image in real time, and directly output high-resolution satellite cloud image through pixel convolution calculation.

[0017] The grid calculation method formula based on CGMS specification is: In the formula, It means that the meteorological satellite points to the center of the earth where the target object is a known quantity. represents the meteorological satellite spin vector, express and The angle between the two can be measured by observing the row number of the center of the earth in the cloud map. represents the cosine function; In step S3, the construction of the pixel convolutional network includes the following steps: S301, the original low-resolution meteorological satellite cloud image is subjected to three convolution layers of different sizes, and the sizes of the convolution kernels used are , , ,The feature mapping area extracted from the satellite cloud images output by each layer has different sizes, and the features of different sizes are merged by splicing to extract distortion features and eliminate compression artifacts and noise; S302, construct three layers of convolution to extract the main features of the meteorological satellite cloud map. In the first layer of convolution, the number of filters is 64 and the convolution kernel size is In the second convolution layer, the number of filters is 32 and the convolution kernel size is In the third convolution layer, the number of filters is 16 and the convolution kernel size is ; S303, fusing the merged features and the extracted main features, and then extracting rich local features as output features by using a residual dense module, where the residual dense module consists of three residual dense blocks; In step S303, the residual dense module is expressed as: In the formula, represents the input features, represents the output features, represents the residual dense block, represents the residual scaling factor, represents the local features extracted by the first residual dense block, represents the local features extracted by the second residual dense block, represents the local features extracted by the third residual dense block; The construction of the pixel convolutional network also includes the following steps: S304, the activation functions of the first and second convolution layers use the ReLU function, and the activation function of the third convolution layer uses the TanH function, and the size of the input meteorological satellite cloud image and the number of feature channels are consistent. The feature map of is the zoom factor; S305: Add the feature map The feature channels are rearranged into This area corresponds to a high-resolution meteorological satellite cloud map This operation converts the original width to Gao Wei And the number of channels is The feature map is directly converted into a width Gao Wei And the number of channels is The feature map of In step S3, the meteorological satellite cloud image super-resolution processing model is: In the formula, represents the operation based on pixel convolutional network, and Respectively represent The parameters of the layer, Indicates The convolution operation of the layer, and They represent low-resolution meteorological satellite cloud images and high-resolution meteorological satellite cloud images respectively. represents the convolutional layer; In step S4, the loss function is used to continuously adjust the parameters to train the meteorological satellite super-resolution processing model. The loss function is expressed as: In the formula, represents the loss function, Represents the convolution kernel and bias parameters to be learned. represents a neural network that generates fractional pixels, and They represent low-resolution meteorological satellite cloud images and high-resolution meteorological satellite cloud images respectively. Indicates the number of parameter adjustments. , Indicates taking the absolute value; A meteorological satellite cloud image super-resolution processing system based on a pixel convolutional network comprises a cloud image acquisition unit 1, a cloud image processing unit 2, a model building unit 3 and a super-resolution processing unit 4; The cloud image acquisition unit 1 is used to collect and process low-resolution meteorological satellite cloud images, obtain low-resolution satellite cloud images from the meteorological big data cloud platform, and update the satellite cloud images every 15 minutes, and process the low-resolution satellite cloud images after collection; The cloud image processing unit 2 analyzes the satellite cloud image into pixel data, so as to facilitate super-resolution processing at the pixel level; Model building unit 3 builds a meteorological satellite cloud image super-resolution processing model based on a pixel convolutional network, selects meteorological satellite cloud images covering multiple weather characteristics of daytime, nighttime, sunny days, and rainy days, compresses them as training data sets, and selects 10% from the uncompressed original meteorological satellite cloud images as a verification data set, and uses the above data sets to perform super-resolution training on the meteorological satellite cloud image super-resolution processing model, so that the refinement and clarity of the meteorological satellite cloud images are significantly improved; The super-resolution processing unit 4 performs feature extraction on the input low-resolution meteorological satellite cloud image in real time, and directly outputs the high-resolution satellite cloud image through pixel convolution calculation; Specifically, the FY4B satellite cloud image was super-resolutioned 4 times using the meteorological satellite cloud image super-resolution processing model. An RTX4090 GPU can super-resolution an image within 1 second, far exceeding the traditional CPU-based algorithm, thus improving the super-resolution processing speed.

[0018] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0019] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A super-resolution processing method for meteorological satellite cloud images based on pixel convolutional networks, characterized in that: The method comprises the following steps: S1. Obtain low-resolution satellite cloud images from the meteorological big data cloud platform, and update the satellite cloud images every 15 minutes. After collection, process the low-resolution satellite cloud images; S2. Use the grid computing method based on the CGMS specification to locate and register the meteorological satellite cloud image of the designated area, convert it into Mercator projection, and finally parse the satellite cloud image into pixel data to facilitate super-resolution processing at the pixel level; S3. Build a meteorological satellite cloud image super-resolution processing model based on pixel convolutional network, select meteorological satellite cloud images covering various weather characteristics such as daytime, nighttime, sunny days and rainy days, compress them as training data sets, and select 10% of the uncompressed original meteorological satellite cloud images as verification data sets. Use the above data sets to perform super-resolution training on the meteorological satellite cloud image super-resolution processing model, so that the refinement and clarity of the meteorological satellite cloud images are significantly improved; S4. Deploy the meteorological satellite cloud image super-resolution processing model on the meteorological big data cloud platform, perform feature extraction on the input low-resolution meteorological satellite cloud image in real time, and directly output high-resolution satellite cloud image through pixel convolution calculation.

2. The method for super-resolution of meteorological satellite cloud images based on pixel convolutional network according to claim 1, characterized in that: The grid calculation method formula based on the CGMS specification is: ； In the formula, It means that the meteorological satellite points to the center of the earth where the target object is a known quantity. represents the meteorological satellite spin vector, express and The angle between the two can be measured by observing the row number of the center of the earth in the cloud map. Represents the cosine function.

3. The method for super-resolution of meteorological satellite cloud images based on pixel convolutional network according to claim 1, characterized in that: In step S3, the construction of the pixel convolutional network includes the following steps: S301, the original low-resolution meteorological satellite cloud image is subjected to three convolution layers of different sizes, and the sizes of the convolution kernels used are , , ,The feature mapping area extracted from the satellite cloud images output by each layer has different sizes, and the features of different sizes are merged by splicing to extract distortion features and eliminate compression artifacts and noise; S302, construct three layers of convolution to extract the main features of the meteorological satellite cloud map. In the first layer of convolution, the number of filters is 64 and the convolution kernel size is In the second convolution layer, the number of filters is 32 and the convolution kernel size is , in the third convolution layer, the number of filters is 16 and the convolution kernel size is ; S303, the merged features and the extracted main features are fused, and then a residual dense module is used to extract rich local features as output features. The residual dense module is composed of three residual dense blocks.

4. The method for super-resolution of meteorological satellite cloud images based on pixel convolutional network according to claim 3 is characterized in that: In step S303, the residual dense module is represented as: ； In the formula, represents the input features, represents the output features, represents the residual dense block, represents the residual scaling factor, represents the local features extracted by the first residual dense block, represents the local features extracted by the second residual dense block, Represents the local features extracted by the third residual dense block.

5. The method for super-resolution of meteorological satellite cloud images based on pixel convolutional network according to claim 4 is characterized in that: The construction of the pixel convolutional network also includes the following steps: S304, the activation functions of the first and second convolution layers use the ReLU function, and the activation function of the third convolution layer uses the TanH function, and the size of the input meteorological satellite cloud image and the number of feature channels are consistent. The feature map of is the zoom factor; S305: Add the feature map The feature channels are rearranged into This area corresponds to a high-resolution meteorological satellite cloud map This operation converts the original width to Gao Wei And the number of channels is The feature map is directly converted into a width Gao Wei And the number of channels is feature map.

6. The method for super-resolution of meteorological satellite cloud images based on pixel convolutional network according to claim 5, characterized in that: The meteorological satellite cloud image super-resolution processing model is: ； In the formula, represents the operation based on pixel convolutional network, and Respectively represent The parameters of the layer, Indicates The convolution operation of the layer, and They represent low-resolution meteorological satellite cloud images and high-resolution meteorological satellite cloud images respectively. Represents a convolutional layer.

7. The method for super-resolution of meteorological satellite cloud images based on pixel convolutional network according to claim 1, characterized in that: In step S3, the loss function is used to continuously adjust the parameters to train the meteorological satellite cloud image super-resolution processing model, and the loss function is expressed as: ； In the formula, represents the loss function, Represents the convolution kernel and bias parameters to be learned. represents a neural network that generates fractional pixels, and They represent low-resolution meteorological satellite cloud images and high-resolution meteorological satellite cloud images respectively. Indicates the number of parameter adjustments. , Indicates taking the absolute value.

8. A meteorological satellite cloud image super-resolution processing system based on a pixel convolutional network, wherein the meteorological satellite cloud image super-resolution processing system is applicable to the meteorological satellite cloud image super-resolution processing method based on a pixel convolutional network according to any one of claims 1 to 7, characterized in that: It includes a cloud image acquisition unit (1), a cloud image processing unit (2), a model building unit (3) and a super-resolution processing unit (4); The cloud image acquisition unit (1) is used to collect and process low-resolution meteorological satellite cloud images, obtain low-resolution satellite cloud images from a meteorological big data cloud platform, and update the satellite cloud images every 15 minutes, and process the low-resolution satellite cloud images after collection; The cloud image processing unit (2) analyzes the satellite cloud image into pixel data, so as to facilitate super-resolution processing at the pixel level; The model building unit (3) builds a meteorological satellite cloud image super-resolution processing model based on a pixel convolutional network, selects meteorological satellite cloud images covering multiple weather characteristics such as daytime, nighttime, sunny days, and rainy days, compresses them as a training data set, and selects 10% of the uncompressed original meteorological satellite cloud images as a verification data set, and uses the above data set to perform super-resolution training on the meteorological satellite cloud image super-resolution processing model; The super-resolution processing unit (4) performs feature extraction on the input low-resolution meteorological satellite cloud image in real time, and directly outputs a high-resolution satellite cloud image through pixel convolution calculation.

Images