Lightweight meteorological downscaling method and device suitable for any multiple, equipment and storage medium

Abstract

The invention provides a lightweight meteorological downscaling method and device suitable for any multiple, equipment and a storage medium. Relates to the technical field of meteorological marine science. The method includes: selecting an interaction target based on a neighborhood configuration; performing feature extraction on the interaction target to obtain features of the interaction target; constructing a graph structure based on the characteristics of the interaction target; based on the graph structure, constructing a high-order graph attention network; a vector cellular automaton is integrated in the high-order map attention network to obtain an HGAT-VCA model, and lightweight meteorological downscaling suitable for any multiple is realized based on the HGAT-VCA model; wherein the vector cellular automaton is used for calculating the total transition probability and determining the land utilization type. According to the method, the expansion of the urban construction land can be simulated and analyzed from a multi-scale and multi-level angle, the complex process of urban expansion and expansion paths under different scenes can be disclosed, and an existing urban expansion theoretical model is enriched.

Description

Technical Field

[0001] This application relates to the field of urban simulation technology, and particularly to a lightweight meteorological downscaling method, device, equipment, and storage medium applicable to any multiple. Background Art

[0002] Currently, traditional methods applicable to lightweight meteorological downscaling for any multiple include statistical regression, system dynamics, cellular automata (CA), multi-agent systems (MAS), etc. Specifically, statistical regression uses historical data to establish statistical relationships and analyze the driving factors of urban expansion (such as population growth, economic indicators, policies, etc.). For example, through multiple linear regression, logistic regression, etc., it quantifies the impact of driving factors on urban expansion and predicts future land use changes based on land type transfer probabilities. It lacks spatial explicit expression and is difficult to simulate complex spatial interactions. System dynamics regards the city as a complex system composed of subsystems such as population, economy, and environment, and analyzes the dynamic relationships between variables through feedback loops. Cellular automata divides space into grid cells and simulates urban expansion based on neighborhood states and transition rules (such as suitability, constraints). Multi-agent systems drive urban expansion from the bottom up by simulating the decision-making behaviors and interactions of "agents" such as residents, governments, and enterprises.

[0003] There is further room for improvement in the accuracy of traditional methods applicable to lightweight meteorological downscaling for any multiple. With the improvement of big data and computing power, machine learning models have been introduced into lightweight meteorological downscaling for any multiple. Machine learning models such as:

[0004] Random forest, support vector machine: used for predicting land conversion probabilities.

[0005] Convolutional neural network (CNN): extracts spatial features in remote sensing images and predicts urban expansion hotspots.

[0006] Generative adversarial network (GAN): generates high-resolution future urban form scenarios.

[0007] The above-mentioned machine learning models have weak interpretability and rely on a large amount of labeled data. Therefore, combining traditional methods with machine learning is one of the current technical means to improve the accuracy of lightweight meteorological downscaling applicable to any multiple. Currently, existing technologies combine CA, MAS, machine learning, and traditional statistical methods to make up for the deficiencies of single models. For example, the CA-MAS coupling model: CA simulates spatial expansion, and MAS simulates the decision-making of agents; CA driven by deep learning: replaces the manually set transition rules with neural networks. However, the combination of vector cellular automata (VCA) and high-order graph attention networks to achieve lightweight meteorological downscaling applicable to any multiple has not been reported in the existing technology. Summary of the Invention

[0008] The present application provides a lightweight meteorological downscaling method, device, equipment and storage medium applicable to any multiple to solve the above technical problems.

[0009] In a first aspect, the present application provides a lightweight meteorological downscaling method applicable to any multiple, including:

[0010] Construct a feature extractor, input a low-resolution image into the feature extractor, and extract low-resolution features.

[0011] Construct an upsampling layer, which, in response to the input low-resolution features, performs upsampling based on an adaptive upsampling operator and outputs a high-resolution feature map.

[0012] Construct a fusion layer, which is used to fuse the high-resolution feature map, high-resolution auxiliary input, and high-resolution elevation data, and adjust the feature channels of the fused features to a set number of channels, and output a reconstructed high-resolution image.

[0013] In a possible design, the feature extractor includes a plurality of frequency feature enhancement blocks, a front convolutional layer, and a rear convolutional layer; wherein, each frequency feature enhancement block includes two convolutional layers and an FEB layer, and the FEB layer includes a convolutional attention module, an instance normalization layer, and an activation layer. The convolutional attention module is used to enhance the feature representation of multi-channel inputs, and the instance normalization layer is used to process the features output by the convolutional attention module, thereby enhancing the stability and generalization of the model.

[0014] In a possible design, the feature extractor extracts low-resolution features from the low-resolution image in the following manner:

[0015] Input the low-resolution image into the front convolutional layer to obtain an original feature map.

[0016] Input the original feature map into the frequency feature enhancement block. Through the convolutional layer in the frequency feature enhancement block, perform feature extraction to obtain a preliminary feature map, and then process it through a convolutional attention module. The convolutional attention module includes channel attention and spatial attention. The channel attention performs global average pooling and global max pooling on the preliminary feature map respectively to obtain two channel description vectors. Subsequently, input the two channel description vectors into a multi-layer perceptron with shared weights to generate channel attention weights. Finally, normalize the channel attention weights through the Sigmoid activation function and multiply them with the preliminary feature map channel by channel to obtain the first feature map. The spatial attention performs average pooling and max pooling on the first feature map in the channel dimension to obtain two spatial feature maps, splice them, generate spatial attention weights through convolution, and finally apply a Sigmoid activation function to normalize the spatial attention weights and multiply them with the first feature map position by position to obtain the second feature map;

[0017] Use an instance normalization layer to normalize the second feature map to obtain a normalized feature map;

[0018] Apply the ReLU function element-wise to the normalized feature map through an activation layer to obtain a low-resolution feature.

[0019] In a possible design, perform upsampling based on an adaptive upsampling operator to output a high-resolution feature map, including:

[0020] The low-resolution feature Pass through a unit convolution to compress the channels to C enc , to obtain a compressed feature map where represents the real number space, C represents the number of input channels, H represents the image height, W represents the image width, and C enc represents the number of output channels;

[0021] Apply lightweight channel attention to the compressed feature map F enc , and the application method is as follows: Use global average pooling to obtain a channel descriptor Generate attention weights through cross-channel interaction Use weighted output to obtain the third feature map F eca = w·F enc , to strengthen the important channel features; Use a unit convolution on the third feature map F eca to generate a dynamic kernel where k up is the upsampling kernel size, and the dynamic kernel K corresponds to the recombination weights of the local region at the spatial position (i, j);

[0022] For the third feature map F ecaExtract a local window according to the spatial position (i, j), and use the dynamic kernel K to perform weighted summation on the features within the window to generate preliminary upsampled features where s is the upsampling factor, and the preliminary upsampled feature F is compressed into a single-channel feature map through unit convolution up ; generate a spatial weight map through Sigmoid with a value range of [0, 1], adjust the feature weights to obtain a high-resolution feature map F final = M·F up .

[0023] In a possible design, the high-resolution auxiliary input is an image obtained by upsampling the high-resolution feature map using the bilinear interpolation method

[0024] In a possible design, the fusion layer includes a splicing layer and a convolutional layer, where the splicing layer is used to fuse the high-resolution feature map, the high-resolution auxiliary input, and the high-resolution elevation data to obtain a fused feature, and the convolutional layer is used to adjust the feature channels of the fused feature to a set number of channels and output a reconstructed high-resolution image

[0025] In a second aspect, the present application provides a lightweight meteorological downscaling device applicable to any multiple, including:

[0026] A feature extraction module configured to construct a feature extractor, input a low-resolution image into the feature extractor, and extract low-resolution features

[0027] An upsampling module configured to construct an upsampling layer, and the upsampling layer performs upsampling based on an adaptive upsampling operator in response to the input low-resolution features and outputs a high-resolution feature map

[0028] A feature fusion module configured to construct a fusion layer, where the fusion layer is used to fuse the high-resolution feature map, the high-resolution auxiliary input, and the high-resolution elevation data, and adjust the feature channels of the fused feature to a set number of channels and output a reconstructed high-resolution image

[0029] In a third aspect, an embodiment of the present application provides an electronic device, including: at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the lightweight meteorological downscaling method applicable to any multiple as described in the first aspect and various possible designs of the first aspect above

[0030] Fourthly, an embodiment of the present application provides a computer-readable storage medium, in which computer-executable instructions are stored. When a processor executes the computer-executable instructions, the lightweight meteorological downscaling method applicable to any multiple as described in the first aspect and various possible designs of the first aspect is implemented.

[0031] Fifthly, an embodiment of the present application provides a computer program product, including a computer program. When the computer program is executed by a processor, the lightweight meteorological downscaling method applicable to any multiple as described in the first aspect and various possible designs of the first aspect is implemented.

[0032] The present application has at least the following beneficial effects:

[0033] 1. A new residual module with a convolutional attention block (CBAM) is proposed to replace the old residual block, and it is stacked as a feature extraction module.

[0034] 2. Batch normalization (BN) is changed to instance normalization (IN) to flexibly handle the case of relatively high input data resolution.

[0035] 3. To solve the "checkerboard artifact" problem and avoid the problems in CARAFE, an adaptive upsampling operator - CSCARAFE is invented, which can identify the importance of each channel and perform flexible upsampling. Description of the Drawings

[0036] The drawings here are incorporated into the description and form a part of this description, showing embodiments consistent with the present application, and are used together with the description to explain the principles of the present application.

[0037] Figure 1 It is the architecture diagram of the lightweight meteorological downscaling model provided by the embodiment of the present application;

[0038] Figure 2 It is the flowchart of a lightweight meteorological downscaling method applicable to any multiple provided by the embodiment of the present application;

[0039] Figure 3 It is the structural schematic diagram of the lightweight meteorological downscaling device applicable to any multiple provided by the embodiment of the present application;

[0040] Figure 4 It is the structural schematic diagram of the electronic device provided by the embodiment of the present application.

[0041] Through the above drawings, the clear embodiments of the present application have been shown, and there will be more detailed descriptions later. These drawings and text descriptions are not intended to limit the scope of the concept of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. Detailed Implementation Modes

[0042] Exemplary embodiments will be described in detail herein, and examples thereof are shown in the accompanying drawings. When the following description refers to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0043] In the technical solution of the present application, the collection, storage, use, processing, transmission, provision, and disclosure of information such as financial data, user data, or medical image data all comply with the provisions of relevant laws and regulations and do not violate public order and good customs.

[0044] It should be noted that in the embodiments of the present application, certain industry-existing solutions such as certain software, components, models, etc. may be mentioned. They should be regarded as exemplary, and their purpose is only to illustrate the feasibility in the implementation of the technical solution of the present application, but it does not mean that the applicant has already or necessarily used this solution.

[0045] It should be noted that in the embodiments of the present application, certain industry-existing solutions such as certain software, components, models, etc. may be mentioned. They should be regarded as exemplary, and their purpose is only to illustrate the feasibility in the implementation of the technical solution of the present application, but it does not mean that the applicant has already or necessarily used this solution.

[0046] The following uses specific embodiments to describe in detail the technical solution of the present application and how the technical solution of the present application solves the above technical problems. These several specific embodiments below can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0047] The embodiments of the present application provide a lightweight meteorological downscaling method applicable to any multiple. As Figure 1 shown, it is the architecture diagram of the lightweight meteorological downscaling model provided by the embodiments of the present application. The lightweight meteorological downscaling method applicable to any multiple can be implemented based on the model as Figure 1 shown. Specifically, please combine Figure 2 shown, which is the flowchart of the lightweight meteorological downscaling method applicable to any multiple provided by the embodiments of the present application. The lightweight meteorological downscaling method applicable to any multiple is implemented by the following steps S10 - S30.

[0048] S10. Construct a feature extractor, input a low-resolution image into the feature extractor, and extract low-resolution features.

[0049] It should be noted that the low-resolution images include, but are not limited to, meteorological maps collected by weather stations.

[0050] Exemplarily, the purpose of step S10 is to extract low-resolution features from the low-resolution images, which can be based on a feature extractor that operates in the low-resolution space. It consists of 16 low-resolution feature enhancement blocks (LR-FEBs), a pre-convolutional layer, and a post-convolutional layer, and is used to enhance the extraction of low-resolution feature maps. Since residual connections are introduced internally, each LR-FEB can be regarded as a mini-residual network. The entire LR-FEB includes two convolutional layers with 3×3 convolutional kernels and a FEB composed of a convolutional attention module CBAM, an instance normalization layer IN, and a relu activation layer. The convolutional attention module can well enhance the feature representation of multi-channel inputs, and the IN layer can well process the features passing through CBAM to enhance the stability and generalization of the model.

[0051] Specifically, the low-resolution image first passes through a pre-convolutional layer to obtain an original feature map (H×W×C). Then, after entering the LR-FEB, it is first extracted by the convolutional layer to obtain a preliminary feature map (H×W×C′). Secondly, it passes through CBAM, which includes channel attention and spatial attention. The channel attention performs global average pooling and global max pooling on the feature map respectively to obtain two channel description vectors (size 1×1×C′ 1×1×C′). Subsequently, the two vectors are input into a multi-layer perceptron (MLP) with shared weights to generate channel attention weights (size 1×1×C′ 1×1×C′). Finally, the channel attention weights are normalized by the Sigmoid activation function and multiplied with the preliminary feature map channel by channel to enhance the feature response of important channels, obtaining the first feature map. The spatial attention performs average pooling and max pooling on the feature map after channel attention (the first feature map) in the channel dimension to obtain two spatial feature maps (size H×W×1 H×W×1). Subsequently, the two are concatenated and a 7×7 convolution is used to generate spatial attention weights (size H×W×1 H×W×1). Finally, a Sigmoid activation function is also applied to normalize the spatial attention weights and multiply them with the first feature map position by position to highlight important spatial regions, obtaining the second feature map.

[0052] The second feature map output after passing through CBAM will enter the IN layer, which normalizes each channel of each sample independently, that is, calculates the mean and variance of a single sample and single channel for standardization. The formula is as follows:

[0053]

[0054] In the formula, x' represents the normalized feature map, x represents the second feature map, μ represents the overall mean, σ 2represents the overall variance, ε represents a very small constant to prevent the denominator from being 0, γ represents a scaling factor, and β represents a translation factor.

[0055] Finally, a ReLU activation layer is used to apply the ReLU function (f(x)=max(0,x)) to the normalized feature map element by element.

[0056] S20. Construct an upsampling layer. The upsampling layer performs upsampling based on an adaptive upsampling operator in response to the input low-resolution features, and outputs a high-resolution feature map.

[0057] In step S20, the upsampling layer is used to map the extracted low-resolution features to the high-resolution space. The upsampling module CSCARAFE used in the model is an adaptive upsampling operator based on the content-aware feature reconstruction operator CARAFE. A lightweight channel attention is inserted into the original kernel prediction branch.

[0058] Specifically, the high-resolution images obtained by CSCARAFE may still exhibit blurred edges or inadequate representation of specific local regions in the high-frequency portions of the climate image. To avoid this, the third part of the method designs a high-resolution feature enhancement module (HR-FEB) to compensate for the shortcomings of CSCARAFE in high-resolution feature representation. This module consists of a convolutional attention block (CBAM), an instance normalization layer (IN) module, and a convolutional layer with a 3×3 kernel; CBAM provides the ability to dynamically enhance features, while IN enhances stability. This layer is finally passed through a convolutional layer to resize the feature channels to make them suitable for the fusion layer.

[0059] The feature map entering the CSCARAFE module first needs to be channel compressed and input low-resolution features After 1x1 convolution, the channel is compressed to C enc , get the compressed feature map Reduce the amount of calculation. Then compress the feature map F enc Apply lightweight channel attention (ECA module), which processes the features as follows: first, global average pooling is performed to obtain channel descriptors Attention weights are then generated through cross-channel interaction (1D convolution) Finally, the weighted output is the third feature map F eca =w·F enc , strengthen the important channel features. After the ECA module, the F eca Generate dynamic kernel using 1x1 convolution where k up is the upsampling kernel size (e.g. 3x3), and kernel K corresponds to the reorganization weight of the local region at the spatial position (i, j).

[0060] Content reorganization extracts a local window (such as a 3x3 neighborhood) from the third feature map F at position (i,j), and uses a dynamic kernel K to weighted sum the features within the window to generate a preliminary upsampled feature. Where s is the upsampling factor. Spatial attention (SA) compresses the preliminary upsampled feature F through a 1x1 convolution up into a single-channel feature map; and then generates a spatial weight map through Sigmoid with a value range of [0,1]. Finally, the feature weights are adjusted to obtain a high-resolution feature map: F final = M·F up , suppressing low-confidence regions (such as background noise) and enhancing key regions (such as object edges).

[0061] S30. Construct a fusion layer, which is used to fuse the high-resolution feature map, high-resolution auxiliary input, and high-resolution elevation data, and adjust the feature channels of the fused features to a set number of channels, and output the reconstructed high-resolution image.

[0062] Exemplarily, the fusion layer can consist of a concatenation layer and a convolutional layer with a 3×3 convolutional kernel. The image obtained by upsampling the input X (i.e., the low-resolution input map) using bilinear interpolation is used as the high-resolution auxiliary input and connected to the concatenation layer. The concatenation part combines the high-resolution feature map, high-resolution auxiliary input, and high-resolution elevation data obtained from the first three parts. The obtained result passes through the last convolutional layer, which does not learn any features and is only used to adjust the feature channels to reconstruct the final output channels. The auxiliary input is introduced into the fusion layer to optimize the reconstruction of the high-resolution image and improve the stability of the network to complete the climate downscaling task on different elements. It should be noted that the digital elevation data is obtained by interpolating the 1km elevation data from the Spatio-Temporal Three Poles Big Data Platform.

[0063] In this embodiment, the method proposed in this application is compared with 4 meteorological downscaling methods proposed in the recent three years and a classic image super-resolution method. In the 2x, 4x, 8x, 10x downscaling of monthly average rainfall and the 4x and 10x downscaling of monthly average 2m temperature experiments, this method obtains the best results in the three metrics of PSNR, MSE, and SSIM, and far exceeds the second-best method.

[0064] The comparison results are shown in Table 1 and Table 2, where EDSR is a classic image super-resolution algorithm, and SRDRN, DeepRU, YNet, and FSRCNN-ESM are excellent meteorological downscaling methods proposed in papers in the recent three years.

[0065] Table 1 Monthly average rainfall downscaling experiment

[0066]

[0067] Table 2 Monthly average 2m air temperature downscaling experiment

[0068]

[0069] Figure 3 The figure is a schematic structural diagram of a lightweight meteorological downscaling device applicable to any multiple provided by an embodiment of the present application. An embodiment of the present application also provides a lightweight meteorological downscaling device applicable to any multiple. As Figure 3 shown, the lightweight meteorological downscaling device applicable to any multiple includes:

[0070] A feature extraction module 301, configured to construct a feature extractor, input a low-resolution image into the feature extractor, and extract low-resolution features;

[0071] An upsampling module 302, configured to construct an upsampling layer, and the upsampling layer upsamples based on an adaptive upsampling operator in response to the input low-resolution features, and outputs a high-resolution feature map;

[0072] A feature fusion module 303, configured to construct a fusion layer, and the fusion layer is used to fuse the high-resolution feature map, high-resolution auxiliary input, and high-resolution elevation data, and adjust the feature channels of the fused features to a set number of channels, and output a reconstructed high-resolution image.

[0073] In some embodiments, the feature extractor includes a plurality of frequency feature enhancement blocks, a front convolutional layer, and a rear convolutional layer; wherein, the frequency feature enhancement block includes two convolutional layers and an FEB layer, and the FEB layer includes a convolutional attention module, an instance normalization layer, and an activation layer, and the convolutional attention module is used to enhance the feature representation of multi-channel inputs, and the instance normalization layer is used to process the features output by the convolutional attention module, thereby enhancing the stability and generalization of the model.

[0074] In some embodiments, the feature extractor extracts low-resolution features from the low-resolution image in the following manner:

[0075] Input the low-resolution image into the front convolutional layer to obtain an original feature map;

[0076] Input the original feature map into the frequency feature enhancement block, extract features through the convolutional layer in the frequency feature enhancement block to obtain a preliminary feature map, and process it through a convolutional attention module. The convolutional attention module includes channel attention and spatial attention. The channel attention performs global average pooling and global max pooling on the preliminary feature map respectively to obtain two channel description vectors. Subsequently, input the two channel description vectors into a multi-layer perceptron with shared weights to generate channel attention weights. Finally, normalize the channel attention weights through the Sigmoid activation function and multiply them with the preliminary feature map channel by channel to obtain the first feature map. The spatial attention performs average pooling and max pooling on the first feature map in the channel dimension, splices the two obtained spatial feature maps and generates spatial attention weights through convolution. Finally, apply a Sigmoid activation function to normalize the spatial attention weights and multiply them with the first feature map position by position to obtain the second feature map;

[0077] Normalize the second feature map using an instance normalization layer to obtain a normalized feature map;

[0078] Apply the ReLU function element-wise to the normalized feature map through an activation layer to obtain a low-resolution feature.

[0079] In some embodiments, perform upsampling based on an adaptive upsampling operator to output a high-resolution feature map, including:

[0080] The low-resolution feature Pass through a unit convolution to compress the channels to C enc to obtain a compressed feature map where represents the real number space, C represents the number of input channels, H represents the image height, W represents the image width, and C enc represents the number of output channels;

[0081] Apply lightweight channel attention to the compressed feature map F enc The application method is: use global average pooling to obtain a channel descriptor Generate attention weights through cross-channel interaction Use weighted output to obtain the third feature map F eca = w·F enc to strengthen the important channel features; for the third feature map F eca Use a unit convolution to generate a dynamic kernel where k up is the upsampling kernel size, and the dynamic kernel K corresponds to the recombination weights of the local region at the spatial position (i, j);

[0082] For the third feature map F ecaExtract a local window according to the spatial position (i, j), and use the dynamic kernel K to perform weighted summation on the features within the window to generate a preliminary upsampled feature where s is the upsampling factor, and the preliminary upsampled feature F is compressed into a single-channel feature map through unit convolution up ; generate a spatial weight map through Sigmoid with a value range of [0, 1], adjust the feature weights, and obtain the high-resolution feature map F final = M·F up .

[0083] In some embodiments, the high-resolution auxiliary input is an image obtained by upsampling the low-resolution input image using the bilinear interpolation method

[0084] In some embodiments, the fusion layer includes a splicing layer and a convolutional layer, where the splicing layer is used to fuse the high-resolution feature map, the high-resolution auxiliary input, and the high-resolution elevation data to obtain a fused feature, and the convolutional layer is used to adjust the feature channels of the fused feature to a set number of channels and output the reconstructed high-resolution image

[0085] The lightweight meteorological downscaling device applicable to any multiple provided by the embodiments of the present application can be used to execute the technical solutions of the lightweight meteorological downscaling method applicable to any multiple in the above embodiments, and its implementation principle and technical effects are similar and will not be elaborated here

[0086] Figure 4 This is a schematic structural diagram of an electronic device provided by an embodiment of the present application. As Figure 4 shown, the electronic device may include: a processor 401 and a memory 402. Among them, the processor 401 and the memory 402 can communicate; exemplarily, the processor 401 and the memory 402 communicate through a communication bus 403

[0087] The processor 401 executes the computer execution instructions stored in the memory 402, so that the processor 401 executes the solutions in the above embodiments. The processor 401 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit ASIC, a field-programmable gate array FPGA, or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components

[0088] The communication bus 403 can be a peripheral component interconnect (PCI) bus, an extended industry standard architecture (EISA) bus, or the like. The system bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of convenience of representation, only a thick line is used in the figure, but it does not mean that there is only one bus or one type of bus. The transceiver is used to implement communication between the database access device and other computers (such as clients, read-write libraries, and read-only libraries). The memory may include random access memory (RAM), and may also include non-volatile memory.

[0089] The electronic device provided by the embodiment of the present application can be the terminal device of the above embodiment.

[0090] The embodiment of the present application also provides a computer-readable storage medium, in which computer instructions are stored. When the computer instructions run on a computer, the computer is enabled to execute the technical solutions of the above embodiments applicable to the lightweight meteorological downscaling method of any multiple.

[0091] The embodiment of the present application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium, and when at least one processor executes the computer program, the technical solutions of the above embodiments applicable to the lightweight meteorological downscaling method of any multiple can be implemented.

[0092] In several embodiments provided by the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the division of modules is only a logical function division. In actual implementation, there may be other division methods. For example, multiple modules can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces, indirect couplings or communication connections of devices or modules, and can be in electrical, mechanical or other forms.

[0093] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they can be located in one place, or can be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to implement the solution of this embodiment.

[0094] In addition, each functional module in various embodiments of the present application may be integrated into a processing unit, may exist physically alone for each module, or two or more modules may be integrated into one unit. The units formed by the above modules may be implemented in the form of hardware, or may be implemented in the form of a combination of hardware and software functional units.

[0095] The integrated modules implemented in the form of software functional modules may be stored in a computer-readable storage medium. The above software functional modules are stored in a storage medium, including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) or a processor to execute some steps of the methods in various embodiments of the present application.

[0096] It should be understood that the above processor may be a central processing unit (CPU for short), or may also be other general-purpose processors, digital signal processors (DSP for short), application specific integrated circuits (ASIC for short), etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the method disclosed in combination with the invention can be directly embodied as being executed and completed by a hardware processor, or can be executed and completed by a combination of hardware and software modules in the processor.

[0097] The memory may include high-speed RAM memory, and may also include non-volatile storage NVM, such as at least one disk memory, and may also be a USB flash drive, a mobile hard disk, a read-only memory, a magnetic disk, or an optical disc, etc.

[0098] The bus may be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, the bus in the drawings of the present application is not limited to only one bus or one type of bus.

[0099] The above storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The storage medium can be any available medium accessible by a general-purpose or special-purpose computer.

[0100] An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can be located in an application specific integrated circuit (ASIC). Of course, the processor and the storage medium can also exist as discrete components in an electronic control unit or a master control device.

[0101] Those of ordinary skill in the art can understand that all or part of the steps to implement the above method embodiments can be completed by hardware related to program instructions. The foregoing program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps including those of the above method embodiments; and the foregoing storage medium includes various media that can store program codes, such as ROM, RAM, magnetic disk, or optical disk.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A lightweight meteorological downscaling method applicable to any multiple, characterized in that, Comprising: Construct a feature extractor, input a low-resolution image into the feature extractor, and extract low-resolution features; Construct an upsampling layer, which, in response to the input low-resolution features, performs upsampling based on an adaptive upsampling operator and outputs a high-resolution feature map; Construct a fusion layer, which is used to fuse the high-resolution feature map, high-resolution auxiliary input, and high-resolution elevation data, and adjust the feature channels of the fused features to a set number of channels, and output a reconstructed high-resolution image.

2. The method according to claim 1, wherein The feature extractor includes a plurality of frequency feature enhancement blocks, a front convolutional layer, and a rear convolutional layer; wherein, the frequency feature enhancement block includes two convolutional layers and an FEB layer, and the FEB layer includes a convolutional attention module, an instance normalization layer, and an activation layer. The convolutional attention module is used to enhance the feature representation of multi-channel inputs, and the instance normalization layer is used to process the features output by the convolutional attention module, thereby enhancing the stability and generalization of the model.

3. The method according to claim 2, characterized in that, The feature extractor extracts low-resolution features from the low-resolution image in the following manner: Input the low-resolution image into the front convolutional layer to obtain an original feature map; Input the original feature map into the frequency feature enhancement block, perform feature extraction through the convolutional layer in the frequency feature enhancement block to obtain a preliminary feature map, and process it through the convolutional attention module. The convolutional attention module includes channel attention and spatial attention. The channel attention performs global average pooling and global max pooling on the preliminary feature map respectively to obtain two channel description vectors, and then inputs the two channel description vectors into a multi-layer perceptron with shared weights to generate channel attention weights. Finally, normalize the channel attention weights through a Sigmoid activation function and multiply them with the preliminary feature map channel by channel to obtain a first feature map. The spatial attention performs average pooling and max pooling on the first feature map in the channel dimension, splices the two spatial feature maps and generates spatial attention weights through convolution, and finally applies a Sigmoid activation function to normalize the spatial attention weights and multiply them with the first feature map position by position to obtain a second feature map; Use the instance normalization layer to normalize the second feature map to obtain a normalized feature map; Apply the ReLU function element-wise to the normalized feature map through the activation layer to obtain low-resolution features.

4. The method according to claim 1, wherein Performing upsampling based on an adaptive upsampling operator and outputting a high-resolution feature map includes: The low-resolution features are compressed to C channels through unit convolution enc to obtain a compressed feature map where represents the real number space, C represents the number of input channels, H represents the image height, W represents the image width, and C enc represents the number of output channels; For the compressed feature map F enc Apply lightweight channel attention in the following way: obtain the channel descriptor using global average pooling Generate attention weights through cross-channel interaction Use the weighted output to obtain the third feature map F eca = w·F enc , strengthening the features of important channels; for the third feature map F eca Use unit convolution to generate a dynamic kernel where k up is the upsampling kernel size, and the dynamic kernel K corresponds to the recombination weights of the local region at the spatial position (i, j); For the third feature map F eca Extract a local window at the spatial position (i, j), and use the dynamic kernel K to perform weighted summation on the features within the window to generate a preliminary upsampled feature where s is the upsampling factor, and the preliminary upsampled feature F is compressed into a single-channel feature map through unit convolution up ; generate a spatial weight map through Sigmoid with a value range of [0, 1], adjust the feature weights to obtain the high-resolution feature map F final = M·F up .

5. The method according to claim 1, characterized in that, The high-resolution auxiliary input is an image obtained by upsampling the low-resolution input image using bilinear interpolation.

6. The method according to claim 1, wherein The fusion layer includes a splicing layer and a convolutional layer. The splicing layer is used to fuse the high-resolution feature map, high-resolution auxiliary input, and high-resolution elevation data to obtain fused features, and the convolutional layer is used to adjust the feature channels of the fused features to a set number of channels and output a reconstructed high-resolution image.

7. A lightweight meteorological downscaling device applicable to any multiple, characterized in that, Comprising: A feature extraction module configured to construct a feature extractor, input a low-resolution image into the feature extractor, and extract low-resolution features; An upsampling module, configured to construct an upsampling layer, which upsamples in response to the input low-resolution features based on an adaptive upsampling operator and outputs a high-resolution feature map; A feature fusion module, configured to construct a fusion layer, which is used to fuse the high-resolution feature map, the high-resolution auxiliary input, and the high-resolution elevation data, and adjust the feature channels of the fused features to a set number of channels, and output a reconstructed high-resolution image.

8. An electronic device, characterized in that, Comprising: A processor, and a memory communicatively connected to the processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory to implement the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, Computer-executable instructions are stored in the computer-readable storage medium, and when the computer-executable instructions are executed by the processor, they are used to implement the method according to any one of claims 1-6.

10. A computer program product, characterized in that, Comprising a computer program, which when executed by the processor implements the method according to any one of claims 1-6.

Images