Method for estimating urban vegetation carbon storage by mixed swin transformer conditional generative adversarial network
By using a hybrid SwinTransformer conditional generative adversarial network combined with multi-source remote sensing image data, high-precision estimation and uncertainty quantification of urban vegetation carbon storage were achieved. This solved the accuracy and generalization problems of remote sensing technology in urban vegetation carbon storage estimation, and improved the reliability and practicality of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIV OF GEOSCIENCES (WUHAN)
- Filing Date
- 2026-03-06
- Publication Date
- 2026-05-29
AI Technical Summary
Existing remote sensing technologies suffer from low accuracy, insufficient generalization ability, and difficulty in quantifying uncertainty in urban vegetation carbon storage estimation, making it difficult to meet the needs of large-scale, high-frequency, and dynamic monitoring.
A hybrid Swing Transformer conditional generative adversarial network is adopted, which combines Landsat8 and PALSAR-2 remote sensing image data. Through adaptive hierarchical feature fusion of CNN local features and Swing Transformer global features, a multi-scale discriminator is constructed. A three-stage training strategy is adopted to generate accurate carbon storage prediction results and provide uncertainty information.
It improves the accuracy and reliability of urban vegetation carbon storage estimation, provides uncertainty information of the estimation results, and enhances the practicality and reliability of the model, making it suitable for urban ecological planning and carbon trading.
Smart Images
Figure CN122116141A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a remote sensing estimation method for urban vegetation carbon storage, specifically a remote sensing estimation method for urban vegetation carbon storage using a hybrid SwingTransformer conditional generative adversarial network, belonging to the field of remote sensing image processing and urban ecological planning technology. Background Technology
[0002] Against the backdrop of the co-evolution of global urbanization and climate change, the accurate quantification of carbon storage in urban vegetation, as the core carrier of carbon cycling in urban ecosystems, has become a key scientific issue supporting sustainable urban development and the implementation of carbon neutrality strategies. Generally speaking, achieving high-precision dynamic monitoring of carbon storage in urban vegetation not only helps to scientifically assess the carbon sequestration capacity and climate regulation function of urban ecosystems, but also provides an important data foundation and decision-making basis for urban green space system planning, ecological asset accounting, carbon trading mechanism design, and climate adaptation governance.
[0003] Traditional carbon storage estimation methods based on field surveys offer advantages such as high accuracy and standardization, but suffer from limitations including high cost, low efficiency, and sparse spatial sampling, making it difficult to meet the needs of large-scale, high-frequency, and dynamic urban monitoring. Specifically, remote sensing technology, with its advantages of macroscopic, rapid, and repeatable observations, has become the primary technical approach for estimating urban vegetation carbon storage. However, current remote sensing-based urban vegetation carbon storage retrieval still faces significant challenges: on the one hand, the highly heterogeneous urban environment and fragmented vegetation distribution make it difficult to construct accurate models linking remote sensing features to carbon storage; on the other hand, existing models often rely on specific scenarios or data sources, lack generalization ability, and the uncertainty of the results is difficult to quantify effectively, severely restricting the application and promotion of remote sensing technology in urban vegetation carbon storage detection.
[0004] In existing technologies: 1) Empirical modeling methods estimate carbon storage by establishing statistical relationships between ground-measured carbon storage data and derived features such as remote sensing spectra / texture. However, these model relationships are highly dependent on specific regional vegetation types, site conditions (such as topography and soil), and climate background, resulting in limited spatial generalization ability. Furthermore, the model accuracy is highly dependent on the representativeness and distribution range of the sample data. If the number of sample plots is insufficient or fails to cover the complete carbon storage gradient, systematic bias is easily introduced. 2) Physical modeling methods are based on the physical relationship between radiative transfer mechanisms and vegetation structure parameters. This method inverts vegetation structure parameters from remote sensing signals and then uses biophysical relationships such as allometric growth equations and carbon content to estimate carbon storage. Despite having a clear mechanistic basis, the model structure is complex, computationally expensive, and sensitive to input parameters. The parameterization process often faces problems such as insufficient ground validation data and scale mismatch, which limits its operational application. 3) Deep learning methods are data-driven and automatically learn the complex nonlinear mapping between remote sensing features and carbon storage through deep neural networks. They do not require manual feature design and have strong potential for information mining and pattern recognition. However, research in this field is still in its early stages. Existing models often have problems such as unstable accuracy and reliance on a large amount of labeled data in carbon storage estimation. In particular, they lack the ability to quantify the uncertainty of prediction results, which affects the reliability and practical applicability of the models. Summary of the Invention
[0005] The purpose of this invention is to provide a remote sensing estimation method for urban vegetation carbon storage using a hybrid SwingTransformer conditional generative adversarial network in order to solve at least one of the above-mentioned technical problems. This method can fully exploit the local and global features in multi-source remote sensing images to improve the accuracy of vegetation carbon storage estimation, while providing uncertainty information of the estimation results to improve the reliability and practicality of the model.
[0006] This invention achieves the above objective through the following technical solution: a remote sensing estimation method for urban vegetation carbon storage using a hybrid Swing Transformer conditional generative adversarial network, comprising the following steps:
[0007] S1. Preprocess Landsat8 and PALSAR-2 remote sensing image data;
[0008] S2. Construct CNN local features and extract CNN local feature branches and construct Swing Transformer global features and extract Swing Transformer global feature branches. The features extracted by the CNN local feature branches and Swing Transformer global feature branches are combined through an adaptive hierarchical feature fusion module. The CNN local feature branches, Swing Transformer global feature branches and the adaptive hierarchical feature fusion module together form the generator of the model.
[0009] S3. Construct a multi-scale discriminator to judge the input remote sensing image patch and the generated carbon storage image based on the spatial scale, as well as the difference and authenticity between the input remote sensing image patch and the real carbon storage image.
[0010] S4. A three-stage training strategy is adopted, including generator pre-training, generator adversarial training, and generator ensemble training, to output the final carbon storage prediction results and uncertainty.
[0011] As a further aspect of the present invention: S1 specifically includes:
[0012] S11. Collect Landsat 8 and PALSAR-2 remote sensing image data, and perform radiometric calibration and edge noise removal on them;
[0013] S12. Cropping, geometric registration, and normalization of Landsat8 and PALSAR-2 remote sensing image data are performed to maintain spatial consistency and data distribution consistency among multi-source remote sensing data, which facilitates feature extraction later.
[0014] As a further aspect of the present invention: In S2, constructing CNN local features and extracting CNN local feature branches specifically includes:
[0015] Construct a Unet network based on CNN;
[0016] The Unet network consists of an encoder, a decoder, and skip connections. The encoder serves as the main encoder and employs a four-layer coding structure, with each layer including a 3×3 convolution, an activation function, and a downsampling operation.
[0017] Based on a four-layer encoding structure, local feature branches of the CNN are extracted layer by layer to obtain local spatial features.
[0018] As a further aspect of the present invention: In S2, constructing the global features of the Swing Transformer and extracting the global feature branches of the Swing Transformer specifically includes:
[0019] The Swin Transformer structure is constructed using a four-layer encoding structure, consisting of four levels of Swin Transformer modules. This structure is responsible for capturing long-range dependencies and forming global feature maps at four scales corresponding to the CNN encoder. These maps are used as global feature extraction paths that run in parallel with the CNN encoder and are embedded into the main Unet framework to form a dual encoder structure.
[0020] As a further aspect of the present invention: In S2, the local features output by the CNN encoder and the global features output by the SwinTransformer encoder are jointly input into the adaptive hierarchical feature fusion module. The global features are enhanced using a channel attention mechanism, and the local features are enhanced using a spatial attention mechanism. Feature fusion is then achieved through a multi-layer neural network. The formula for the adaptive hierarchical feature fusion module is as follows:
[0021]
[0022] in, Represents global features. To represent local features, AvgPool represents average pooling, and MaxPool represents max pooling. This indicates an element-wise multiplication operation. This represents a convolution operation with a kernel size of n×n. This represents the Sigmoid function, and GN represents the batch normalization operation. Indicates the final characteristics after fusion;
[0023] After four encoding stages, the fused multi-scale features are input into a four-layer decoder. The decoder contains a four-level upsampling structure and makes skip connections with the four scale features corresponding to the dual encoders, respectively, to restore the spatial resolution step by step, and finally obtain the predicted image of urban vegetation carbon storage.
[0024] As a further aspect of the present invention: S3 specifically includes:
[0025] S31. The original image with resolution (H,W) is progressively reduced by convolutional downsampling to obtain images of three scales: the original scale (H,W), the first downsampling scale (H / 2,W / 2), and the second downsampling scale (H / 4,W / 4). The images of the three scales are used to represent the overall structure, regional texture, and local details, respectively.
[0026] S32. The discriminator with the same structure is used for images of all three scales. Each discriminator consists of five 5×5 convolutional layers. Multi-scale texture features and spatial patterns are extracted through layer-by-layer convolution and downsampling to achieve the discrimination of the authenticity features of the input carbon storage image.
[0027] S33. Input the images at the three scales mentioned above into the corresponding discriminators to obtain the authenticity discrimination results at the three scales. Then, perform a weighted sum of the three discrimination outputs to obtain the final comprehensive authenticity discrimination result. The objective function of the multi-scale discriminator is shown below:
[0028]
[0029] in, The values represent the weights of the authenticity judgment results at each scale, with D1, D2, and D3 representing the discriminators corresponding to the three scales of the image, respectively. The discriminator is D. n The objective function of a Wasserstein conditional generative adversarial network with gradient penalty at time (n=1,2,3).
[0030] As a further aspect of the present invention: S4 specifically includes:
[0031] S41. Generator Pre-training Phase: To obtain a suitable generator initialization and accelerate the training process, the generator is first pre-trained with mean squared error (MSE) loss. The MSE calculation formula is as follows:
[0032]
[0033] Among them, y i Represents the actual value. Indicates the predicted value;
[0034] S42. Generative Adversarial Training Phase: The generator and discriminator are trained simultaneously. Multi-source remote sensing images are input into the generator to produce carbon storage images. The discriminator receives real carbon storage images or generated carbon storage images to determine their authenticity. The generator and discriminator are continuously optimized through adversarial training, so that the generated carbon storage images closely resemble the real carbon storage images in terms of spatial pattern and numerical distribution.
[0035] To address the stability issue during training of conditional generative adversarial networks (GANs), a Wasserstein conditional GAN model with gradient penalty is introduced, combined with L1 loss to improve the accuracy of carbon storage inversion. The specific loss function formula is as follows:
[0036]
[0037] Among them, y i Represents the actual value. Indicates the predicted value. Let denote the objective function of the Wasserstein conditional generative adversarial network with gradient penalty, l represent the conditional information, x represent the true carbon storage map, and z represent the random noise vector. This indicates the generation of a carbon storage image. Indicates the interpolated sample. Indicates the discriminator output For interpolated samples The gradient of G* is given by G*, which represents the composite loss function of the generator. This represents the weight of the generative adversarial network loss in the generator composite loss;
[0038] S43. In the generator ensemble training phase, while continuing generative adversarial training, a generator network ensemble is constructed through the generator ensemble strategy to enhance the stability and generalization ability of the model.
[0039] S44. Use the trained generator set to generate multiple carbon storage images, calculate the mean of the generated carbon storage images to obtain the final carbon storage prediction result, and calculate the standard deviation as an evaluation of the prediction uncertainty.
[0040] The beneficial effects of this invention are:
[0041] 1) The data input of this invention uses open-source remote sensing images. It improves the accuracy of carbon storage estimation through a hybrid network architecture. At the same time, it outputs the uncertainty information of the estimation through the built-in probability modeling module of the model, which overcomes the dependence on the preset parameterization form, fills the shortcoming of the existing technology that only provides "single-point numerical results", improves the reliability and practicality of the model, and provides support for urban ecological planning and carbon sink trading.
[0042] 2) This invention designs a parallel dual-encoder generator structure that integrates CNN and Swin Transformer, which efficiently captures local features and global semantic information representations at different scales. At the same time, an adaptive hierarchical feature fusion module is designed to effectively fuse semantic information between features at different scales in each branch, thereby improving the network's representation learning ability.
[0043] 3) This invention introduces a multi-scale discriminator to achieve coordinated optimization of global semantic structure consistency and local fine information authenticity through a division of labor and cooperation mode. By using cross-scale conditional input to suppress semantic shifts in the generation process, it improves training stability and enhances the accuracy of carbon storage inversion. Attached Figure Description
[0044] Figure 1 This is a flowchart of the model training technology of the present invention;
[0045] Figure 2 This is an architecture diagram of the generator in the hybrid Swing Transformer conditional generative adversarial network of this invention;
[0046] Figure 3 This is an architecture diagram of the discriminator in the hybrid Swing Transformer conditional generative adversarial network of this invention. Detailed Implementation
[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0048] Example 1, as Figures 1 to 3 As shown in the figure, this embodiment provides a remote sensing estimation method for urban vegetation carbon storage using a hybrid Swing Transformer conditional generative adversarial network. This method includes the following steps:
[0049] S1. Preprocess Landsat8 and PALSAR-2 remote sensing image data;
[0050] S2. Construct CNN local features and extract CNN local feature branches and construct Swing Transformer global features and extract Swing Transformer global feature branches. The features extracted by the CNN local feature branches and Swing Transformer global feature branches are combined through an adaptive hierarchical feature fusion module. The CNN local feature branches, Swing Transformer global feature branches and the adaptive hierarchical feature fusion module together form the generator of the model, thereby realizing the generation of carbon storage images that take into account the multi-scale fusion of local and global features of remote sensing images.
[0051] S3. Construct a multi-scale discriminator to judge the input remote sensing image patch and the generated carbon storage image from multiple spatial scales, as well as the difference and authenticity between the input remote sensing image patch and the real carbon storage image.
[0052] S4. A three-stage training strategy is adopted, including a generator pre-training stage, a generator adversarial training stage, and a generator ensemble training stage, to output the final carbon storage prediction results and uncertainty.
[0053] Example 2, in addition to all the technical features included in Example 1, also includes:
[0054] S1 specifically includes:
[0055] S11. Collect Landsat 8 and PALSAR-2 remote sensing image data, and perform radiometric calibration and edge noise removal on them;
[0056] S12. Cropping, geometric registration, and normalization of Landsat8 and PALSAR-2 remote sensing image data are performed to maintain spatial consistency and data distribution consistency among multi-source remote sensing data, which facilitates feature extraction later.
[0057] In S2, constructing CNN local features and extracting CNN local feature branches specifically includes:
[0058] Construct a Unet network based on CNN;
[0059] The Unet network consists of an encoder, a decoder, and skip connections. The encoder, as the main encoder, adopts a four-layer coding structure, with each layer including a 3×3 convolution, an activation function, and a downsampling operation.
[0060] Based on a four-layer coding structure, local feature branches of the CNN are extracted layer by layer to obtain local spatial features.
[0061] Constructing the global features of the Swin Transformer and extracting the global feature branches of the Swin Transformer specifically includes:
[0062] The Swin Transformer structure is constructed using a four-layer encoding structure, consisting of four levels of Swin Transformer modules. This structure is responsible for capturing long-range dependencies and forming global feature maps at four scales corresponding to the CNN encoder. These maps are used as global feature extraction paths that run in parallel with the CNN encoder and are embedded into the main Unet framework to form a dual encoder structure.
[0063] The local features output from the CNN encoder and the global features output from the Swin Transformer encoder are jointly input into the adaptive hierarchical feature fusion module. The global features are enhanced using a channel attention mechanism, while the local features are enhanced using a spatial attention mechanism. Feature fusion is then achieved through a multi-layer neural network. The formula for the adaptive hierarchical feature fusion module is as follows:
[0064]
[0065] in, Represents global features. To represent local features, AvgPool represents average pooling, and MaxPool represents max pooling. This indicates an element-wise multiplication operation. This represents a convolution operation with a kernel size of n×n. This represents the Sigmoid function, and GN represents the batch normalization operation. Indicates the final characteristics after fusion;
[0066] After four encoding stages, the fused multi-scale features are input into a four-layer decoder. The decoder contains a four-level upsampling structure and makes skip connections with the four scale features corresponding to the dual encoders, respectively, to restore the spatial resolution step by step, and finally obtain the predicted image of urban vegetation carbon storage.
[0067] S3 specifically includes:
[0068] S31. The original image with resolution (H,W) is progressively reduced through convolutional downsampling to obtain images at three scales: the original scale (H,W), the first downsampling scale (H / 2,W / 2), and the second downsampling scale (H / 4,W / 4). The images at these three scales are used to represent the overall structure, regional texture, and local details, respectively.
[0069] S32. The discriminator with the same structure is used for images of all three scales. Each discriminator consists of five 5×5 convolutional layers. Multi-scale texture features and spatial patterns are extracted through layer-by-layer convolution and downsampling to achieve the discrimination of the authenticity features of the input carbon storage image.
[0070] S33. Input the images at the three scales mentioned above into the corresponding discriminators to obtain the authenticity discrimination results at the three scales. Then, perform a weighted sum of the three discrimination outputs to obtain the final comprehensive authenticity discrimination result. The objective function of the multi-scale discriminator is shown below:
[0071]
[0072] in, The values represent the weights of the authenticity judgment results at each scale, with D1, D2, and D3 representing the discriminators corresponding to the three scales of the image, respectively. The discriminator is D. n The objective function of a Wasserstein conditional generative adversarial network with gradient penalty at time (n=1,2,3).
[0073] S4 specifically includes:
[0074] S41. Generator Pre-training Phase: To obtain a suitable generator initialization and accelerate the training process, the generator is first pre-trained with mean squared error (MSE) loss. The MSE calculation formula is as follows:
[0075]
[0076] Among them, y i Represents the actual value. Indicates the predicted value;
[0077] S42. Generative Adversarial Training Phase: The generator and discriminator are trained simultaneously. Multi-source remote sensing images are input into the generator to produce carbon storage images. The discriminator receives real carbon storage images or generated carbon storage images to determine their authenticity. The generator and discriminator are continuously optimized through adversarial training, so that the generated carbon storage images closely resemble the real carbon storage images in terms of spatial pattern and numerical distribution.
[0078] To address the stability issue during training of conditional generative adversarial networks (GANs), a Wasserstein conditional GAN model with gradient penalty is introduced, combined with L1 loss to improve the accuracy of carbon storage inversion. The specific loss function formula is as follows:
[0079]
[0080] Among them, y i Represents the actual value. Indicates the predicted value. Let denote the objective function of the Wasserstein conditional generative adversarial network with gradient penalty, l represent the conditional information, x represent the true carbon storage map, and z represent the random noise vector. This indicates the generation of a carbon storage image. Indicates the interpolated sample. Indicates the discriminator output For interpolated samples The gradient of G* is given by G*, which represents the composite loss function of the generator. This represents the weight of the generative adversarial network loss in the generator composite loss;
[0081] S43. In the generator ensemble training phase, while continuing generative adversarial training, a generator network ensemble is constructed through the generator ensemble strategy to enhance the stability and generalization ability of the model.
[0082] S44. Use the trained generator set to generate multiple carbon storage images, calculate the mean of the generated carbon storage images to obtain the final carbon storage prediction result, and calculate the standard deviation as an evaluation of the prediction uncertainty.
[0083] Working Principle: Based on a conditional generative adversarial network (GAN) framework, a hybrid generator integrating local features and global semantics is constructed. Its core utilizes a convolutional neural network to extract local detail textures from remote sensing images, while simultaneously employing a SwinTransformer to model global long-range dependencies. An adaptive attention fusion module then deeply integrates these two types of features to form a comprehensive deep representation of the spatial pattern of vegetation carbon storage. A multi-scale discriminator identifies the authenticity of the generated results at the global, regional, and local levels, driving the generator to produce carbon storage maps that are both structurally sound and detail-accurate. The model is progressively optimized through three stages: generator pre-training, adversarial training, and ensemble training. Finally, the ensemble model generates multiple sets of prediction results, and the average is used as the final carbon storage estimation map. The standard deviation is calculated as a pixel-level quantitative output of uncertainty, thereby achieving high-precision and high-reliability remote sensing estimation of urban vegetation carbon storage.
[0084] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
[0085] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A remote sensing estimation method for urban vegetation carbon storage using a hybrid Swing Transformer conditional generative adversarial network, characterized in that, The remote sensing estimation method for urban vegetation carbon storage includes the following steps: S1. Preprocess Landsat8 and PALSAR-2 remote sensing image data; S2. Construct CNN local features and extract CNN local feature branches and construct Swing Transformer global features and extract Swing Transformer global feature branches. The features extracted by the CNN local feature branches and Swing Transformer global feature branches are processed through an adaptive hierarchical feature fusion module. The CNN local feature branches, Swing Transformer global feature branches and the adaptive hierarchical feature fusion module together form the generator of the model. S3. Construct a multi-scale discriminator to judge the input remote sensing image patch and the generated carbon storage image based on the spatial scale, as well as the difference and authenticity between the input remote sensing image patch and the real carbon storage image. S4. A three-stage training strategy is adopted, including a generator pre-training stage, a generator adversarial training stage, and a generator ensemble training stage, to output the final carbon storage prediction results and uncertainty.
2. The remote sensing estimation method for urban vegetation carbon storage according to claim 1, characterized in that, S1 specifically includes: S11. Collect Landsat 8 and PALSAR-2 remote sensing image data, and perform radiometric calibration and edge noise removal on them; S12. The Landsat8 and PALSAR-2 remote sensing image data are cropped, geometrically registered, and normalized.
3. The remote sensing estimation method for urban vegetation carbon storage according to claim 1, characterized in that, In step S2, constructing CNN local features and extracting CNN local feature branches specifically includes: Construct a Unet network based on CNN; The Unet network consists of an encoder, a decoder, and skip connections. The encoder, as the main encoder, adopts a four-layer coding structure, with each layer including a 3×3 convolution, an activation function, and a downsampling operation. Based on a four-layer encoding structure, local feature branches of the CNN are extracted layer by layer.
4. The remote sensing estimation method for urban vegetation carbon storage according to claim 3, characterized in that, In step S2, constructing the global features of the Swing Transformer and extracting the global feature branches of the Swing Transformer specifically includes: The Swin Transformer structure is constructed using a four-layer encoding structure, consisting of four levels of Swin Transformer modules. This structure is responsible for capturing long-range dependencies and forming global feature maps at four scales corresponding to the CNN encoder. These maps are used as global feature extraction paths that run in parallel with the CNN encoder and are embedded into the main Unet framework to form a dual encoder structure.
5. The remote sensing estimation method for urban vegetation carbon storage according to claim 4, characterized in that, In step S2, the local features output by the CNN encoder and the global features output by the Swin Transformer encoder are jointly input into the adaptive hierarchical feature fusion module. The global features are enhanced using a channel attention mechanism, and the local features are enhanced using a spatial attention mechanism. Feature fusion is then achieved through a multi-layer neural network. The formula for the adaptive hierarchical feature fusion module is as follows: in, Represents global features. To represent local features, AvgPool represents average pooling, and MaxPool represents max pooling. This indicates an element-wise multiplication operation. This represents a convolution operation with a kernel size of n×n. This represents the Sigmoid function, and GN represents the batch normalization operation. Indicates the final characteristics after fusion; After four encoding stages, the fused multi-scale features are input into a four-layer decoder. The decoder contains a four-level upsampling structure and makes skip connections with the four scale features corresponding to the dual encoders, respectively, to restore the spatial resolution step by step, and finally obtain the predicted image of urban vegetation carbon storage.
6. The remote sensing estimation method for urban vegetation carbon storage according to claim 1, characterized in that, S3 specifically includes: S31. The original image with resolution (H,W) is progressively reduced by convolutional downsampling to obtain images of three scales: the original scale (H,W), the first downsampling scale (H / 2,W / 2), and the second downsampling scale (H / 4,W / 4). The images of the three scales are used to represent the overall structure, regional texture, and local details, respectively. S32. The discriminator with the same structure is used for images of all three scales. Each discriminator consists of five 5×5 convolutional layers. Multi-scale texture features and spatial patterns are extracted through layer-by-layer convolution and downsampling to achieve the discrimination of the authenticity features of the input carbon storage image. S33. Input the images at the three scales mentioned above into the corresponding discriminators to obtain the authenticity discrimination results at the three scales. Then, perform a weighted sum of the three discrimination outputs to obtain the final comprehensive authenticity discrimination result. The objective function of the multi-scale discriminator is shown below: in, The values represent the weights of the authenticity judgment results at each scale, with D1, D2, and D3 representing the discriminators for the images at the three scales, respectively. The discriminator is D. n The objective function of a Wasserstein conditional generative adversarial network with gradient penalty at time (n=1,2,3).
7. The remote sensing estimation method for urban vegetation carbon storage according to claim 1, characterized in that, S4 specifically includes: S41. Generator Pre-training Stage: The generator is pre-trained under mean squared error loss, where the MSE calculation formula is as follows: Among them, y i Represents the actual value. Indicates the predicted value; S42. Generative Adversarial Training Phase: The generator and discriminator are trained simultaneously. Multi-source remote sensing images are input into the generator to produce carbon storage images. The discriminator receives real carbon storage images or generated carbon storage images to determine their authenticity. The generator and discriminator are continuously optimized through adversarial training, so that the generated carbon storage images closely resemble the real carbon storage images in terms of spatial pattern and numerical distribution. To address the stability issue during training of conditional generative adversarial networks (GANs), a Wasserstein conditional GAN model with gradient penalty is introduced, combined with L1 loss to improve the accuracy of carbon storage inversion. The specific loss function formula is as follows: Among them, y i Represents the actual value. Indicates the predicted value. Let denote the objective function of the Wasserstein conditional generative adversarial network with gradient penalty, l represent the conditional information, x represent the true carbon storage map, and z represent the random noise vector. This indicates the generation of a carbon storage image. Indicates the interpolated sample. Indicates the discriminator output For interpolated samples The gradient of G* is given by G*, which represents the composite loss function of the generator. This represents the weight of the generative adversarial network loss in the generator composite loss; S43. Generator Ensemble Training Phase: While continuing generative adversarial training, a generator network ensemble is constructed through a generator ensemble strategy to enhance the model's stability and generalization ability. S44. Generate carbon storage images using the trained generator set, calculate the mean of the generated carbon storage images to obtain the final carbon storage prediction result, and calculate the standard deviation as an evaluation of the prediction uncertainty.