Carbon concentration reconstruction method based on physical manifold and multi-scale generative network

Abstract

This application provides a carbon concentration reconstruction method based on physical manifolds and multi-scale generative networks, belonging to the field of carbon monitoring technology. The method includes: real-time acquisition of satellite data and carbon concentration point-line data; encoding the satellite data into a macroscopic background feature map, and extracting node representation vectors from the carbon concentration point-line data using a graph neural network; performing graph-grid cross-modal feature fusion using grid points in the macroscopic background feature map as queries and node representation vectors as keys and values ​​to obtain a high-resolution feature map; inputting the high-resolution feature map into a trained generative adversarial network, which uses wind-guided deformable convolutional layers for feature extraction and sub-pixel upsampling reconstruction, outputting a reconstructed high-resolution carbon concentration field. By introducing physical manifold guidance and multi-scale generative networks, this application achieves accurate fusion and high-fidelity reconstruction of multi-source heterogeneous carbon data from both satellite and ground sources, significantly improving the accuracy and physical consistency of carbon emission monitoring.

Description

Technical Field

[0001] This application belongs to the field of carbon monitoring technology, specifically relating to a carbon concentration reconstruction method based on physical manifolds and multi-scale generative networks. Background Technology

[0002] Against the backdrop of global efforts to address climate change, the precise monitoring and quantitative accounting of greenhouse gas emissions has become a core requirement for ecological and environmental regulation. Currently, greenhouse gas monitoring methods mainly fall into two categories: one is area-source remote sensing observation, primarily based on satellite remote sensing, which offers the advantage of full-area coverage but suffers from drawbacks such as low spatial resolution and significant cloud interference; the other is localized, precise observation, mainly using high-precision ground-based automatic monitoring stations and drones / mobile monitoring vehicles, which offers extremely high accuracy but very limited spatial coverage. Therefore, how to heterogeneously assimilate discrete, high-precision "point-line" measured data with wide-area, low-resolution "area" satellite observation data to reconstruct a high-fidelity carbon concentration field that combines wide-area coverage and high resolution has become a critical technical bottleneck that urgently needs to be overcome in the field of atmospheric environmental monitoring.

[0003] For spatial fusion and high-resolution reconstruction of concentration fields from multi-source carbon monitoring data, the industry currently offers three main similar implementation schemes. The first is spatial interpolation methods based on traditional geostatistics, such as Kriging interpolation or inverse distance weighting. These methods acquire discrete carbon concentration data from ground monitoring stations or mobile monitoring vehicles, perform pure mathematical fitting and deduction based on a function of geospatial distance, and generate a continuous carbon concentration distribution map. The second is deep learning image super-resolution algorithms based on traditional computer vision, such as standard convolutional neural networks or generative adversarial networks. These methods equate low-resolution satellite carbon concentration grids to low-resolution digital images, utilize the network to learn the pixel mapping relationship from low resolution to high resolution, and then overlay ground observation data onto the image through simple grid alignment. The third is traditional atmospheric transport models and data assimilation algorithms, such as the WRF-Chem model. These methods describe the transport and diffusion processes of pollutants by solving complex differential equations.

[0004] However, the aforementioned existing technical solutions have significant drawbacks in practical applications. First, traditional deep learning super-resolution algorithms use regular square convolutional kernels for isotropic feature extraction. However, carbon emission plumes are highly directional due to wind field influences, resulting in reconstructed plume edges often exhibiting aerodynamically incompatible circular or irregular artifacts, leading to severe physical topological distortion. Second, ground station data is a discrete non-Euclidean spatial point cloud, while satellite data is a continuous Euclidean spatial grid. Existing technologies typically use simple spatial interpolation to grid ground points before inputting them into the network. This not only introduces significant interpolation errors but also causes high-precision measured data to be smoothed out during rasterization, failing to fulfill its "anchor point" calibration function and resulting in fragmented multi-scale heterogeneous data fusion. Finally, conventional AI reconstruction models only optimize pixel-level visual similarity loss, leading to a discrepancy between the local integral sum of the reconstructed high-resolution concentration field and the total carbon mass observed by the original satellite pixels, violating the laws of physical conservation. Furthermore, traditional physical assimilation models require solving extremely complex differential equations and adjoint matrices, resulting in extremely low computational efficiency and making it difficult to meet real-time monitoring requirements. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this application provides a carbon concentration reconstruction method based on physical manifolds and multi-scale generative networks. The technical problem to be solved by this application is achieved through the following technical solution: A carbon concentration reconstruction method based on physical manifolds and multi-scale generative networks includes: S1: Real-time acquisition of satellite data and carbon concentration point / line data; S2: Encode the satellite data into a macroscopic background feature map and use a graph neural network to extract node representation vectors from the carbon concentration dot-line data; S3: Using the grid points in the macro background feature map as queries and the node representation vectors as keys and values, perform graph-grid cross-modal feature fusion to obtain a high-resolution feature map; S4: Input the high-resolution feature map into the trained generative adversarial network, so that the trained generative adversarial network uses wind-guided deformable convolutional layers for feature extraction and sub-pixel upsampling reconstruction, and outputs a reconstructed high-resolution carbon concentration field; wherein, the wind-guided deformable convolutional layer generates the offset of each convolutional kernel sampling point, so that the receptive field of the convolutional kernel conforms to the physical diffusion manifold of the carbon emission plume.

[0006] Beneficial effects: (1) This application employs a wind-guided deformable convolutional layer, which generates the offset of the convolutional kernel sampling point based on the wind direction and wind speed vector of each pixel. This stretches the sampling point in the downwind direction and compresses it in the upwind direction, so that the receptive field of the convolutional kernel conforms to the physical diffusion manifold of the carbon emission plume. Compared with traditional generative networks that use regular square convolutional kernels, the high carbon concentration region reconstructed in this application can perfectly fit the hydrodynamic diffusion pattern of the real atmosphere, eliminate circular or irregular artifacts that violate aerodynamics, and significantly improve physical fidelity.

[0007] (2) This application abandons the traditional approach of spatially interpolating discrete point clouds before rasterizing them. Instead, it directly constructs a graph structure from the high-precision measured carbon concentration point and line data obtained from ground monitoring stations and UAVs / mobile monitoring vehicles. It uses a graph neural network to extract node representation vectors and performs feature fusion through a graph-grid cross-modal attention mechanism, using grid points of satellite feature maps as queries and node representation vectors as keys and values. This scheme avoids the secondary error introduced by the interpolation process, maximizes the utilization of the absolute accuracy of small ground sample data, and achieves sub-pixel-level accurate spatial registration and feature anchoring of point, line, and surface data in the hidden layer of the network.

[0008] (3) This application introduces macroscopic mass conservation loss and convective gradient penalty loss into the training loss function of the generative adversarial network. The mass conservation loss forces the sum of carbon concentration integrals in the reconstructed high-resolution carbon concentration field to be equal to the total carbon mass observed in the original low-resolution satellite pixel region. The convective gradient penalty loss constrains the gradient tensor of the reconstructed concentration field to be consistent with the wind field direction based on the measured wind field vector, thus suppressing the false topology of upwind diffusion. The above constraints enable the refined carbon emission data reconstructed by AI (artificial intelligence) to be consistent with the large-scale satellite observations in terms of total amount, and have the legal traceability capability as national-level carbon emission regulatory data.

[0009] (4) This application adopts a generative adversarial network with offline physical training and online fast inference. During the online operation phase, only a small number of matrix multiplication operations are required to output a high-resolution carbon concentration field in milliseconds. The computational efficiency is orders of magnitude higher than that of traditional physical assimilation models based on solving differential equations (such as WRF-Chem), which can meet the needs of online second-level early warning and dynamic monitoring of sudden greenhouse gas leaks.

[0010] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating a carbon concentration reconstruction method based on physical manifolds and multi-scale generative networks provided in this application. Figure 2 This is a schematic diagram of carbon concentration reconstruction based on physical manifolds and multi-scale generative networks provided in this application; Figure 3 This is a detailed schematic diagram of each step in the carbon concentration reconstruction process provided in this application. Detailed Implementation

[0012] The present application will be described in further detail below with reference to specific embodiments, but the implementation of the present application is not limited thereto.

[0013] Combination Figures 1 to 3 , Figure 1 This is a flowchart illustrating a carbon concentration reconstruction method based on physical manifolds and multi-scale generative networks provided in this application. Figure 2 This is a schematic diagram of carbon concentration reconstruction based on physical manifolds and multi-scale generative networks provided in this application; Figure 3 This is a detailed schematic diagram of each step in the carbon concentration reconstruction provided in this application. This application provides a carbon concentration reconstruction method based on physical manifolds and multi-scale generative networks, including: S1: Real-time acquisition of satellite data and carbon concentration point-line data; wherein, the satellite data includes low-resolution carbon column concentration grid data, wind field vector data, and digital elevation model data; the carbon concentration point-line data includes discrete point carbon concentration data acquired by ground monitoring stations, and trajectory point carbon concentration data collected by UAVs or mobile monitoring vehicles along the moving trajectory.

[0014] S2: Encode the satellite data into a macroscopic background feature map and use a graph neural network to extract node representation vectors from the carbon concentration dot-line data; In one specific embodiment of this application, S2 includes: S21: Encode the satellite data to generate a macroscopic background feature map; In one specific embodiment of this application, S21 includes: S211: The low-resolution carbon column density grid data, wind field vector data, and digital elevation model data in the satellite data are stitched together along the channel dimension to obtain the stitched data; S212: Input the spliced ​​data into the encoder branch of the multi-scale cascaded architecture to extract multi-scale macroscopic background feature maps.

[0015] refer to Figure 2 and Figure 3This application stitches together the low-resolution carbon column density grid from the Sentinel-5P / TROPOMI satellite with wind field vectors (wind speed, wind direction) and a digital elevation model (DEM), and extracts macroscopic background feature maps through a multi-scale cascaded architecture (such as the encoder branch of U-Net++). TROPOMI, the only payload on the Sentinel-5P satellite, stands for Tropospheric Monitoring Instrument. It is an imaging spectrometer that inverts the concentrations of various trace gases and aerosols in the atmosphere by measuring the sunlight reflected and radiated by the Earth.

[0016] S22: A graph neural network is used to construct a graph structure and extract features from the carbon concentration point and line data to generate node representation vectors.

[0017] In one specific embodiment of this application, S22 includes: S221: Construct a graph structure from the acquired discrete point carbon concentration data and / or trajectory point carbon concentration data, where the graph node features are the measured carbon concentration values, and the edge weights between graph nodes are determined by the physical spatial distance between nodes and the wind field connectivity. This application can collect carbon concentration data from discrete points at ground monitoring stations or from trajectory points from drones or mobile monitoring vehicles.

[0018] S222: A graph neural network is used to propagate and update the features of the graph structure, and node representation vectors containing high-frequency spatial details are extracted.

[0019] refer to Figure 2 and Figure 3 This application abandons traditional spatial interpolation and constructs a graph structure from 1D (one-dimensional) ground station and 2D (two-dimensional) UAV trajectory points. The node features are high-precision measured concentrations, and the edge weights are determined by the physical distance between nodes and the wind field connectivity. A graph neural network (GNN) is used to extract node representation vectors containing high-frequency details.

[0020] This application can also construct a coordinate-based multilayer perceptron (MLP). Using satellite grid features as conditional encoding, and inputting the continuous physical coordinates (x, y, z) of the ground station / UAV into the MLP, the continuous concentration field at those coordinates can be directly predicted. Employing an MLP can replace schemes using NeRF (Neural Radiation Field) or continuous coordinate mapping techniques, and can also replace schemes using graph structures for sub-pixel fusion.

[0021] S3: Using the grid points in the macro background feature map as queries and the node representation vectors as keys and values, perform graph-grid cross-modal feature fusion to obtain a high-resolution feature map; In one specific embodiment of this application, S3 includes: S31: Use the feature vector of each grid point in the macro background feature map as the query vector; S32: Use the node representation vectors extracted by the graph neural network as key vectors and value vectors; S33: Calculate the attention weights of the query vector and the key vector using the cross-attention mechanism, and perform a weighted summation of the value vectors based on the attention weights to obtain a high-resolution feature map.

[0022] refer to Figure 2 and Figure 3 To achieve accurate registration of point / line data with area data, this application introduces a cross-attention mechanism from the Transformer architecture. The feature map grid points extracted from satellites serve as the query, while the ground / mobile node representations extracted by the GNN serve as the key and value. This cross-attention mechanism adaptively "extracts" density details from high-precision ground nodes based on wind field similarity and spatial distance, injecting them into the low-resolution satellite feature map to achieve sub-pixel-level feature anchoring.

[0023] After the point cloud data is meshed, this application can directly use STN (Spatial Deformation Network) to calculate the affine transformation matrix (including translation, scaling, and rotation) between the ground station mesh and the satellite mesh, and achieve spatial alignment and fusion of the two through matrix multiplication.

[0024] S4: Input the high-resolution feature map into the trained generative adversarial network, so that the trained generative adversarial network uses wind-guided deformable convolutional layers for feature extraction and sub-pixel upsampling reconstruction, and outputs a reconstructed high-resolution carbon concentration field; wherein, the wind-guided deformable convolutional layer generates the offset of each convolutional kernel sampling point, so that the receptive field of the convolutional kernel conforms to the physical diffusion manifold of the carbon emission plume.

[0025] In this application, the trained generative adversarial network generates the offset of each convolutional kernel sampling point using wind-guided deformable convolutional layers; the sampling points are stretched in the downwind direction and compressed in the upwind direction using the offsets to obtain deformed sampling points; the deformed sampling points are used to perform feature sampling on the high-resolution feature map, and then sub-pixel upsampling reconstruction is performed to output a reconstructed high-resolution carbon concentration field.

[0026] In this application, a wind-guided deformable convolutional layer is used to obtain the sampling points of each convolutional kernel on the high-resolution feature map, as well as the corresponding wind direction and wind speed vectors. Based on the wind direction and wind speed vectors, the offset of each convolutional kernel sampling point relative to the standard sampling grid is calculated, causing the sampling points to be stretched in the downwind direction and compressed in the upwind direction, resulting in deformed sampling points. The wind-guided deformable convolutional layer ultimately generates the offset of each convolutional kernel sampling point.

[0027] Get the convolution kernel The standard sampling position of each sampling point relative to the center of convolution. and the wind field vector corresponding to the convolution center. ;in, This represents the wind field vector corresponding to the convolution center. and The superscript indicates the components of the wind field vector on the horizontal and vertical coordinate axes, respectively. T Represents the transpose of a matrix or vector; Calculate the wind speed based on the wind field vector. and unit wind direction vector And determine the unit crosswind vector based on the unit wind direction vector. ; In the formula, This represents the wind speed calculated based on the wind field vector. This represents a preset positive number used to prevent the denominator from being zero. Represents a unit direction vector along the wind direction; and These represent the components of the unit wind direction vector on the horizontal and vertical coordinate axes, respectively. Represents the unit lateral wind vector orthogonal to the unit wind direction vector; The standard sampling locations are projected onto the unit wind direction vector and the unit crosswind direction vector, respectively, to obtain the downwind component. and crosswind weight ; In the formula, Indicates the standard sampling position Unit wind direction vector The projection on, i.e., the first The windward component of each sampling point; Indicates the standard sampling position Unit lateral wind vector The projection on, i.e., the first Crosswind component at each sampling point; Calculate the scaling factor based on the wind speed. and in accordance with Calculate the first The offset of each convolutional kernel sampling point relative to the standard sampling grid; In the formula, This represents the scaling factor of the convolution kernel sampling grid, determined based on wind speed. This represents the preset maximum scaling factor, used to limit the maximum impact of wind speed on the deformation of the sampling grid. This indicates a preset reference wind speed, used to adjust the rate at which the expansion coefficient changes with the wind speed; express The reciprocal of is used to control the degree of compression of the sampling grid in the crosswind direction; Indicates the first The offset vector of each convolution kernel sampling point relative to the standard sampling grid; Then we get: ; In the formula, Indicates the first The deformed sampling position after applying the offset to the sampling points of each convolution kernel.

[0028] This application can also use a standard square two-dimensional convolution kernel, but instead of directly convolving the feature map, it treats the wind field vector field as an equivalent optical flow field, uses grid resampling technology to pre-distort or deform the feature map in the windward direction, then performs regular convolution on the distorted feature map, and finally recovers it in reverse.

[0029] This application utilizes the deformed sampling points to perform feature sampling on the high-resolution feature map, and performs weighted summation with the convolution kernel weights to obtain the deformed convolutional feature map; the deformed convolutional feature map is then reconstructed by sub-pixel upsampling to enlarge it to a high-resolution space, thereby outputting a reconstructed high-resolution carbon concentration field.

[0030] This application allows for the design of an auxiliary "super network" without altering the geometric sampling point positions of the convolution kernels. Wind speed and direction are used as inputs to the super network, and the one-dimensional asymmetric convolution kernels required for the backbone network (such as...) are calculated and generated in real time. or The weights of the convolution kernel.

[0031] refer to Figure 3This application innovatively introduces Wind-Guided Deformable Convolution (WG-DCN) into the generator of a Generative Adversarial Network (GAN), abandoning the standard two-dimensional convolutional kernel. This wind-guided deformable convolution layer uses the wind direction and speed vector of each pixel as the offset learning condition for the deformable convolution. This causes the sampling points of the convolutional kernel to be stretched in the downwind direction and compressed in the upwind direction. Thus, when the generator performs sub-pixel upsampling reconstruction, its feature receptive field completely conforms to the physical diffusion manifold of carbon emission plumes in the atmosphere, resulting in smoother edges and aerodynamically more efficient reconstructed carbon plume edges.

[0032] In one specific embodiment of this application, reference is made to Figure 2 and Figure 3 The trained generative adversarial network is obtained through training using a joint loss function, which includes adversarial loss, structural similarity loss, macroscopic mass conservation loss, and convection gradient penalty loss; the joint loss function is expressed as: ; in, To combat the losses; For structural similarity loss; As a macroscopic mass conservation loss, in the high-resolution carbon concentration field obtained by its forced constraint reconstruction, the sum of the carbon concentration integrals corresponding to the original low-resolution satellite pixel region is equal to the total carbon mass observed in that low-resolution pixel. As a convection gradient penalty loss, the gradient tensor of the high-resolution carbon concentration field reconstructed based on the measured wind field vector constraint is consistent with the wind field direction; , , These are the corresponding weighting coefficients.

[0033] In the high-resolution concentration field map generated by the macroscopic mass conservation loss constraint in this application, the sum of the concentration integrals corresponding to a certain low-resolution satellite original pixel region must be equal to the observed value of that low-resolution original pixel. The convective gradient penalty loss calculates the theoretical spatial gradient of the concentration based on the measured wind field, constraining the gradient tensor of the generated image to be consistent with the direction of the physical wind field, preventing the appearance of false topology due to backwind diffusion.

[0034] The training process of the generative adversarial network (GAN) in this application can refer to relevant technologies, all of which use a training set for multiple iterations of training. The GAN consists of a generator and a discriminator. The generator is responsible for generating data, and the discriminator supervises the generator's generation results. This application uses a joint loss function to adjust the generator's parameters. The joint loss function of this application combines macroscopic mass conservation and gradient joint constraints, enabling the refined carbon emission data reconstructed by AI to be consistent with large-scale satellite observations in terms of total amount, and possessing the legal traceability capability as national-level carbon emission regulatory data.

[0035] It is worth noting that the terms "first" and "second" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0036] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of this application and should not be construed as limiting the specific implementation of this application to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of this application, and all such modifications or substitutions should be considered within the scope of protection of this application.

Claims

1. A carbon concentration reconstruction method based on physical manifolds and multi-scale generative networks, characterized in that, include: S1: Real-time acquisition of satellite data and carbon concentration point / line data; S2: Encode the satellite data into a macroscopic background feature map and use a graph neural network to extract node representation vectors from the carbon concentration dot-line data; S3: Using the grid points in the macro background feature map as queries and the node representation vectors as keys and values, perform graph-grid cross-modal feature fusion to obtain a high-resolution feature map; S4: Input the high-resolution feature map into the trained generative adversarial network, so that the trained generative adversarial network uses wind-guided deformable convolutional layers for feature extraction and sub-pixel upsampling reconstruction, and outputs a reconstructed high-resolution carbon concentration field; wherein, the wind-guided deformable convolutional layer generates the offset of each convolutional kernel sampling point, so that the receptive field of the convolutional kernel conforms to the physical diffusion manifold of the carbon emission plume.

2. The carbon concentration reconstruction method based on physical manifolds and multi-scale generative networks according to claim 1, characterized in that, The satellite data includes low-resolution carbon column concentration grid data, wind field vector data, and digital elevation model data; the carbon concentration point and line data includes discrete point carbon concentration data acquired by ground monitoring stations, as well as trajectory point carbon concentration data collected by UAVs or mobile monitoring vehicles along their movement trajectories.

3. The carbon concentration reconstruction method based on physical manifolds and multi-scale generative networks according to claim 2, characterized in that, S2 include: S21: Encode the satellite data to generate a macroscopic background feature map; S22: A graph neural network is used to construct a graph structure and extract features from the carbon concentration point and line data to generate node representation vectors.

4. The carbon concentration reconstruction method based on physical manifolds and multi-scale generative networks according to claim 3, characterized in that, S21 includes: S211: The low-resolution carbon column density grid data, wind field vector data, and digital elevation model data in the satellite data are stitched together along the channel dimension to obtain the stitched data; S212: Input the spliced ​​data into the encoder branch of the multi-scale cascaded architecture to extract multi-scale macroscopic background feature maps.

5. The carbon concentration reconstruction method based on physical manifolds and multi-scale generative networks according to claim 3, characterized in that, S22 includes: S221: Construct a graph structure from the acquired discrete point carbon concentration data and / or trajectory point carbon concentration data, where the graph node features are the measured carbon concentration values, and the edge weights between graph nodes are determined by the physical spatial distance between nodes and the wind field connectivity. S222: A graph neural network is used to propagate and update the features of the graph structure, and node representation vectors containing high-frequency spatial details are extracted.

6. The carbon concentration reconstruction method based on physical manifolds and multi-scale generative networks according to claim 1, characterized in that, S3 include: S31: Use the feature vector of each grid point in the macro background feature map as the query vector; S32: Use the node representation vectors extracted by the graph neural network as key vectors and value vectors; S33: Calculate the attention weights of the query vector and the key vector using the cross-attention mechanism, and perform a weighted summation of the value vectors based on the attention weights to obtain a high-resolution feature map.

7. The carbon concentration reconstruction method based on physical manifolds and multi-scale generative networks according to claim 1, characterized in that, S4 include: The high-resolution feature map is input into the trained generative adversarial network (GAN), so that the trained GAN performs the following process: The offset of each convolution kernel sampling point is generated using a wind-guided deformable convolutional layer; The sampling point is stretched in the downwind direction and compressed in the upwind direction using the offset to obtain the deformed sampling point; The deformed sampling points are used to perform feature sampling on the high-resolution feature map, and then sub-pixel upsampling reconstruction is performed to output the reconstructed high-resolution carbon concentration field.

8. The carbon concentration reconstruction method based on physical manifolds and multi-scale generative networks according to claim 7, characterized in that, The offset of each convolution kernel sampling point generated by the wind-guided deformable convolutional layer includes: Obtain the sampling point of each convolutional kernel on the high-resolution feature map, as well as the wind direction and wind speed vector corresponding to that sampling point; The offset of each convolution kernel sampling point relative to the standard sampling grid is calculated based on the wind direction and wind speed vectors, and is expressed as follows: ； in, Indicates the convolution kernel number 1 The offset of each sampling point relative to the standard sampling grid. This represents the scaling factor of the convolution kernel sampling grid, determined based on wind speed. Indicates the standard sampling position Unit wind direction vector Projection on; express The reciprocal of is used to control the degree of compression of the sampling grid in the crosswind direction; This represents the unit lateral wind vector that is orthogonal to the unit wind direction vector.

9. The carbon concentration reconstruction method based on physical manifolds and multi-scale generative networks according to claim 7, characterized in that, The step of using the deformed sampling points to perform feature sampling on the high-resolution feature map, followed by sub-pixel upsampling reconstruction to output a reconstructed high-resolution carbon concentration field includes: The deformed sampling points are used to sample features on the high-resolution feature map, and the results are weighted and summed with the convolution kernel weights to obtain the deformed convolutional feature map. The feature map after deformed convolution is reconstructed by subpixel upsampling to enlarge it to a high-resolution space, thereby outputting a reconstructed high-resolution carbon concentration field.

10. The carbon concentration reconstruction method based on physical manifolds and multi-scale generative networks according to claim 1, characterized in that, The trained generative adversarial network is obtained through training using a joint loss function, which includes adversarial loss, structural similarity loss, macroscopic mass conservation loss, and convection gradient penalty loss; the joint loss function is expressed as: ； in, To combat the losses; For structural similarity loss; As a macroscopic mass conservation loss, in the high-resolution carbon concentration field obtained by its forced constraint reconstruction, the sum of the carbon concentration integrals corresponding to the original low-resolution satellite pixel region is equal to the total carbon mass observed in that low-resolution pixel. As a convection gradient penalty loss, the gradient tensor of the high-resolution carbon concentration field reconstructed based on the measured wind field vector constraint is consistent with the wind field direction; , , These are the corresponding weighting coefficients.

Images