Urban carbon reserve estimation method based on fine-grained source domain alignment and causal constraint
By employing fine-grained source domain alignment and causal constraints, the problems of accuracy and ecological mechanism constraints in cross-regional transfer learning were solved, enabling accurate estimation of urban carbon storage.
Patent Information
- Application Number
- CN202511751176.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-26
- Publication Date
- 2026-02-06
AI Technical Summary
In existing technologies, the accuracy of cross-regional transfer learning and the lack of ecological mechanism constraints in carbon density prediction models lead to the failure to utilize the heterogeneity of the geographical environment within the source domain and the model learning feature weights that contradict ecological principles.
We employ a fine-grained source domain screening and weighted alignment method based on geographical similarity, combined with a self-supervised hierarchical carbon density prediction method based on causal constraints. This method avoids negative transfer through fine-grained source domain alignment and ensures that the model learns the correct factor hierarchy relationships through causal constraints.
It enables automatic identification of land use types from remote sensing images, accurate prediction of carbon density, generation of spatial distribution maps of urban carbon storage, avoids negative migration, and conforms to ecological mechanisms.
Smart Images
Figure CN121481814A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of ecological environment monitoring and carbon storage accounting, and particularly relates to a city carbon storage estimation method based on fine-grained source domain alignment and causal constraint, which is used for accurately quantifying spatial distribution characteristics of city ecological system carbon storage. BACKGROUND
[0002] The city carbon storage estimation method based on remote sensing technology faces two key technical bottlenecks.
[0003] The first technical bottleneck is the precision problem of cross-region transfer learning. The existing domain adaptation method aligns the entire source domain with the target domain as a whole, ignoring the geographical environment heterogeneity within the source domain. Specifically, a city usually contains plain areas, hilly areas, mountainous areas and other geographical environment types. When the dominant geographical environment types of the source domain city and the target domain city are different, the overall alignment strategy will cause serious negative transfer problem. In fact, some local areas of the source domain city may be very similar to the geographical environment of the target domain city, and the knowledge of these local areas is more valuable and easier to transfer, but the existing method cannot automatically identify and utilize this local similarity.
[0004] The second technical bottleneck is the problem that the carbon density prediction model lacks ecological mechanism constraint. Ecological studies show that the explanatory rate of climate factors to carbon density variation usually reaches 60% to 70%, which is the main controlling factor; the explanatory rate of terrain factors is about 20% to 30%; the explanatory rate of human activity factors is about 5% to 10%. However, the existing machine learning method usually inputs all environmental factors equally for end-to-end training, without reflecting this hierarchical decreasing relationship. More seriously, the deep learning model is easy to learn the feature weight allocation that is contrary to the ecological law in the optimization process, and the existing method lacks an effective mechanism to explicitly constrain and verify the factor importance learned by the model.
[0005] Therefore, it is urgent to design a new technical scheme to realize fine-grained alignment of the source domain to fully utilize the local geographical environment similarity, and to ensure that the carbon density prediction model conforms to the ecological hierarchical law through an explicit causal constraint mechanism. SUMMARY
[0006] In view of the above defects of the prior art, the technical problem to be solved by the present application is to provide a city carbon storage estimation method based on fine-grained source domain alignment and causal constraint, which can overcome the negative transfer problem caused by the heterogeneity within the source domain through a fine-grained source domain screening and weighted alignment strategy based on geographical environment similarity, and ensure that the model learns the correct factor hierarchical relationship through a self-supervised hierarchical carbon density prediction method based on causal constraint.
[0007] In order to achieve the above object, the application provides a city carbon storage estimation method based on fine-grained source domain alignment and causal constraint, characterized by comprising the following steps:
[0008] Step one, fine-grained source domain screening based on geographical environment similarity is adopted;
[0009] Step two, cross-regional land use classification is carried out by using a weighted alignment method;
[0010] Step three, a self-supervised hierarchical carbon density prediction method based on causal constraint is used to estimate carbon density;
[0011] Step four, the spatial distribution of city carbon storage is calculated and the result is output.
[0012] Further, step one specifically comprises:
[0013] A0, a domain alignment model trained by source domain city data identifies land use types from target domain remote sensing images;
[0014] A1, geographical environment feature vectors g p =[DEM p , Slope p , NDVI p , LST p ] are extracted for the source domain and the target domain;
[0015] A2, clustering is carried out based on geographical environment features for each source domain, and the source domain is divided into several sub-regions;
[0016] A3, the maximum mean difference distance between the target domain and each source domain sub-region is calculated
[0017] Further, step two specifically comprises:
[0018] B1, alignment weight is calculated based on distance and high-similarity sub-regions are screened;
[0019] B2, a fine-grained weighted domain alignment network is constructed and trained;
[0020] B3, a target domain land use classification map is generated.
[0021] Further, A1 and A2 in step one specifically comprise:
[0022] C1, geographical environment features of source domain pixels are extracted, including elevation DEM p , terrain slope Slope p , normalized vegetation index NDVI p and land surface temperature LST pThe normalized vegetation index is calculated according to the formula:
[0023]
[0024] wherein NIR p is the reflectivity in the near-infrared band, Red p is the reflectivity in the red band;
[0025] C2, Z-score standardization is performed on the feature vectors to eliminate the dimensional differences of different features;
[0026] C3, a K-means clustering algorithm is used, and the number of clusters M is adaptively determined by a silhouette coefficient:
[0027]
[0028] wherein a p is the average distance of sample p to the same class of samples, and b p is the average distance of sample p to the nearest different class. The value of the silhouette coefficient ranges from -1 to 1, and the larger the value, the better the clustering effect. Different cluster numbers are tried, and the cluster number that maximizes the average silhouette coefficient is selected as the optimal cluster number;
[0029] C4, the source domain sub-region division
[0030] Further, B1 and B2 in step two specifically include:
[0031] D1, a Gaussian kernel is used to calculate the maximum mean difference distance:
[0032]
[0033] wherein the kernel bandwidth γ is determined by a median heuristic method: randomly sample sample pairs, and calculate the median of the distance as the kernel bandwidth;
[0034] D2, the alignment weight is calculated, and a softmax function is used:
[0035]
[0036] wherein the temperature parameter τ is set as the standard deviation of all distances, so that when the distance distribution is concentrated, the weight is slightly concentrated, and when the distance distribution is discrete, the weight distribution is smooth;
[0037] D3, high-similarity sub-regions are screened, and the third quartile Q3 of the distance is calculated as the threshold. For sub-regions with a distance , the sub-regions are excluded from the twenty-fifth percentile of the sub-regions with the largest distance;
[0038] D4. Construct a domain alignment network. The feature extractor uses ResNet-50, and the classifier outputs six land use probabilities. Construct an independent domain discriminator for each sub-region.
[0039] D5. Constructing a fine-grained weighted domain alignment loss:
[0040]
[0041] in For source domain sub-regions Losses in the confrontation with the target domain;
[0042] D6. The total loss function combines classification loss and domain alignment loss:
[0043]
[0044] The balance coefficient λ is determined through validation set optimization. Different values are tried on the validation set, and the value of λ that maximizes the classification accuracy is selected.
[0045] Furthermore, step three specifically includes:
[0046] E0, Construct a carbon density training set based on field plot measurement data;
[0047] E1. Extract three layers of environmental factors: climate factor X1 (average annual temperature, annual precipitation, annual solar radiation), topographic factor X2 (altitude, slope, aspect), and anthropogenic factor X3 (population density, nighttime light index, distance from city center).
[0048] E2. Construct a hierarchical carbon density prediction model C = f1(X1) × f2(X2) + f3(X3);
[0049] E3. Construct a self-supervised learning task for factor importance;
[0050] E4. Design a causal constraint loss function;
[0051] E5. Jointly train and output carbon density estimation results.
[0052] Furthermore, steps three, specifically E3 and E4, include:
[0053] F1. For each training sample, keeping other factors constant, apply a small perturbation to the primary climate factor, setting the perturbation amplitude to 10% of the factor's standard deviation, and calculate the change in carbon density ΔC1. Perform the same operation on other factors to obtain ΔC2 and ΔC3. A perturbation amplitude of 10% is sufficient to observe changes in carbon density without exceeding a reasonable range.
[0054] F2. Calculate the true factor importance vector:
[0055]
[0056] This vector represents the percentage contribution of each layer factor to the change in carbon density;
[0057] F3. Add a factor importance prediction branch to the hierarchical prediction model, and output the predicted factor importance vector I. pred ;
[0058] F4. Define self-supervised loss to measure the difference between predicted importance and true importance:
[0059]
[0060] F5. Define causal constraint loss:
[0061]
[0062] The threshold θ is determined by analyzing the statistical distribution of the true importance of climate factors in the training set, and the percentile is taken. A penalty is applied when the predicted importance of a climate factor falls below the threshold.
[0063] F6. Construct the total loss function:
[0064]
[0065] in This represents the mean squared error loss for carbon density prediction. The balance coefficients β1 and β2 were determined through validation set optimization.
[0066] The beneficial effects of this invention are: it can automatically identify land use types from remote sensing images, predict carbon density based on environmental factors, avoid negative migration through fine-grained source domain alignment, ensure ecological mechanisms through causal constraints, and generate a spatial distribution map of carbon storage. Attached Figure Description
[0067] Figure 1 This is a schematic diagram of the overall process of an embodiment of the present invention;
[0068] Figure 2 This is a schematic diagram of a fine-grained source domain screening and weighted alignment process based on geographical environment similarity;
[0069] Figure 3 This is a schematic diagram of a self-supervised hierarchical carbon density prediction process based on causal constraints. Detailed Implementation
[0070] The present invention will be further described below with reference to the accompanying drawings and embodiments:
[0071] This embodiment takes the Yangtze River Delta urban agglomeration as the research object, identifies land use types and estimates the spatial distribution of carbon storage from remote sensing images.
[0072] like Figure 1 As shown, this embodiment provides a method for estimating urban carbon storage based on fine-grained source domain alignment and causal constraints, including the following steps:
[0073] Step 1: Integrate Landsat-8 remote sensing imagery data from three source cities and five target cities. Label the source datasets with land use, creating a labeled dataset containing six categories: forest, grassland, cultivated land, water bodies, construction land, and unused land. Perform cross-regional land use classification using a fine-grained source domain selection and weighted alignment method based on geographic similarity.
[0074] Step 2: Input the land use classification map obtained in Step 1 and the three-layer environmental factor data extracted from data sources such as WorldClim, SRTM DEM, and LandScan into the self-supervised hierarchical carbon density prediction model based on causal constraints, and output the carbon density estimation results.
[0075] Step 3: Combining the land use classification map and carbon density prediction results, calculate the carbon storage of each pixel, generate and output the spatial distribution map of urban carbon storage.
[0076] Furthermore, such as Figure 2 The example shown illustrates fine-grained source domain filtering and weighted alignment based on geographic environment similarity. Step one includes:
[0077] A0. Data preparation is carried out based on three source cities, including a coastal plain city (source S1), an inland hilly city (source S2), and a mountain city (source S3).
[0078] A1. Remote sensing image preprocessing is performed on the source and target domain datasets, including radiometric calibration, atmospheric correction, and geometric correction;
[0079] A2. Divide the source domain dataset into training set, validation set and test set in an 8:1:1 ratio;
[0080] A3. Extract geographical environmental features from all pixels;
[0081] A4. Cluster each source domain based on geographical environmental characteristics;
[0082] A5. Calculate the geographical similarity between the target domain and each source domain sub-region;
[0083] A6. Calculate alignment weights based on similarity and construct a fine-grained weighted domain alignment network.
[0084] Furthermore, the specific steps for extracting geographic environmental features from all pixels as described in A3 are as follows:
[0085] B0. Geographic environmental feature extraction is the foundation of the first core innovation of this invention. Geographic environmental features reflect the geographic environmental conditions of the pixel rather than the spectral response characteristics of the ground object itself.
[0086] B1. Extracting elevation features (DEM) from the SRTM digital elevation model. p Elevation reflects the vertical position of a pixel and is a key indicator for distinguishing plains (elevation less than 100 meters), hills (elevation between 100 and 500 meters), and mountains (elevation greater than 500 meters).
[0087] B2. Calculate the terrain slope feature. p The maximum gradient method using a 3x3 neighborhood window is employed. The slope is measured in degrees and ranges from 0 to 90 degrees. Areas with steep slopes are typically mountains or hills, while areas with gentle slopes are typically plains.
[0088] B3. Calculate the Normalized Difference Vegetation Index (NDVI) p Calculated from the near-infrared and red bands of Landsat-8 imagery:
[0089]
[0090] The NDVI value ranges from [-1, 1]. In vegetation-covered areas, the NDVI is usually greater than 0.2, and in dense vegetation, the NDVI is greater than 0.6.
[0091] B4. Retrieving Land Surface Temperature Characteristics (LST) from the Thermal Infrared Band of Landsat-8 p A single-window algorithm was used to retrieve land surface temperature.
[0092] B5. Combining to form a geographical environment feature vector g p =[DEM p Slope p NDVI p LST p ].
[0093] Furthermore, the specific steps for clustering each source domain based on geographical environmental characteristics as described in A4 are as follows:
[0094] C0. Fine-grained partitioning of the source domain is a key step in the first core innovation of this invention, which divides the source domain into sub-regions with similar geographical environments through clustering.
[0095] C1. The geographical environment feature vector is Z-score standardized. The mean and standard deviation of each feature dimension on the training set are calculated. The standardization process eliminates the dimensional differences of different physical quantities such as altitude (meters), slope (degrees), NDVI (dimensionless), and surface temperature (Kelvin).
[0096] C2. The K-means clustering algorithm is used, which minimizes the sum of squares within each cluster through iterative optimization.
[0097] C3. Determine the optimal number of clusters using the silhouette coefficient, which is defined as:
[0098]
[0099] Where a p Let b be the average Euclidean distance from sample p to other samples of the same class. p Let M be the average distance from sample p to the nearest outlier sample. The silhouette coefficient ranges from -1 to 1; the closer the value is to 1, the better the clustering effect. Try the number of clusters M∈{2,3,4,5,6,7,8}, calculate the average silhouette coefficient, and select the M that maximizes it as the optimal number of clusters.
[0100] C4. Perform K-means clustering to obtain the source domain S. i Sub-regional division. For example, source region S1 (coastal plain city) may be divided into sub-regions such as low-altitude plain area, coastal wetland area, low mountain and hill area, high-density built-up area, and farmland area.
[0101] Furthermore, the specific steps for calculating similarity and constructing a fine-grained weighted domain alignment network as described in A5 and A6 are as follows:
[0102] D0. Calculating the similarity of geographical environment is the core step of the first core innovation of this invention.
[0103] D1. Extract the geographic environment feature vector from the target domain T, and standardize it using the mean and standard deviation calculated on the source domain training set to ensure that the source domain and the target domain are on the same scale.
[0104] D2. The Maximum Mean Difference (MMD) distance is used to measure the difference between distributions, as this distance effectively measures the difference between two distributions. A Gaussian kernel function is chosen.
[0105]
[0106] Where γ is the kernel bandwidth parameter. The kernel bandwidth is determined using a median heuristic: one thousand pairs of samples are randomly sampled from all sub-regions of the source domain and the target domain, and the Euclidean distance between them is calculated. The median of the distance is used as the kernel bandwidth. This method can adapt to the scale of the data.
[0107] D3. Calculate the MMD distance for all source domain sub-regions and the target domain to obtain the distance set;
[0108] D4. Calculate alignment weights based on MMD distance using the softmax function:
[0109]
[0110] The temperature parameter τ is set to the standard deviation of all distances. This setting ensures that when the distance distribution is concentrated, the weights are slightly concentrated in the most similar sub-regions, and when the distance distribution is discrete, the weight distribution is smoother.
[0111] D5. Filter high similarity sub-regions, calculate the third quartile (Q3, 75th percentile) of the distance as the threshold, set the weight to zero for sub-regions with a distance greater than the threshold, and exclude the sub-regions with the largest distance in the top 25%.
[0112] D6. Normalize non-zero weights: Renormalize the weights of the remaining sub-regions so that the sum of all non-zero weights is one.
[0113] D7. Construct a domain alignment network, including a feature extractor (using ResNet-50, outputting 2048-dimensional features), a classifier (outputting six types of land use probabilities), and a domain discriminator (constructing an independent binary discriminator for each source domain sub-region).
[0114] D8. Define the classification loss and use the cross-entropy loss function;
[0115] D9. Define the domain alignment loss. For each source domain sub-region, connect a gradient inversion layer (GRL) before the domain discriminator. The gradient remains unchanged during forward propagation and is inverted during backward propagation to achieve adversarial training.
[0116] D10. Construct a fine-grained weighted domain alignment loss and use adaptive weights to sum multiple adversarial losses:
[0117]
[0118] In this way, sub-regions with high similarity receive greater weight, while sub-regions with large environmental differences are excluded.
[0119] D11, Total Loss Function:
[0120]
[0121] The balancing coefficient λ is selected by trying different values on the validation set to achieve the highest overall classification accuracy on the validation set. In this embodiment, the optimal λ = 0.5.
[0122] D12. Adopt an adversarial training strategy to alternately optimize the feature extractor, classifier, and domain discriminator, and use the trained model to generate a land use classification map of the target domain.
[0123] Furthermore, such as Figure 3 The following is an analysis of self-supervised hierarchical carbon density prediction based on causal constraints. Step two includes:
[0124] F0. Collect field plot data, measure aboveground biomass, belowground biomass, soil organic carbon, and litter carbon using standard sample methods, and calculate carbon density. A total of 300 plots were collected.
[0125] F1. Divide the sample plot data into a training set (210 samples), a validation set (60 samples), and a test set (30 samples) in a ratio of 7:2:1.
[0126] F2. Extract three levels of environmental factors: primary climate factors include annual average temperature, annual precipitation, and annual solar radiation; secondary topographic factors include altitude, slope, and aspect; and tertiary anthropogenic factors include population density, nighttime light index, and distance from the city center.
[0127] F3. Perform Z-score standardization on environmental factors;
[0128] F4. Construct a hierarchical carbon density prediction model;
[0129] F5, Construct a self-supervised learning task for factor importance;
[0130] F6. Design a causal constraint loss function;
[0131] F7. Use a multi-stage training strategy to train the model.
[0132] Furthermore, the specific steps for constructing the hierarchical carbon density prediction model described in F4 are as follows:
[0133] G0, the prediction of stratified carbon density is the basis of the second core innovation of this invention;
[0134] G1. Construct a first-level model f1(X1), predict the basic carbon density based on climate factors, and use the gradient boosting tree GBT algorithm, which achieves strong nonlinear fitting ability by integrating multiple decision trees.
[0135] G2. Construct a secondary model f2(X2), predict correction coefficients based on terrain factors, and employ the Random Forest (RF) algorithm. The range of correction coefficients is constrained to 0.7 to 1.3, reflecting that the moderating effect of terrain factors on carbon density is typically within ±30%.
[0136] G3. Construct a three-level model f3(X3), predict the perturbation term based on human factors, and use the XGBoost algorithm;
[0137] G4. Final carbon density calculation formula:
[0138] C = f1(X1) × f2(X2) + f3(X3)
[0139] The base carbon density multiplied by the correction factor reflects the proportional adjustment effect of topographic factors, while the addition of a perturbation term reflects the external disturbance effect of human activities. This hierarchical structure explicitly encodes the hierarchical dependence of factors.
[0140] Furthermore, the specific steps for constructing self-supervised learning of factor importance and designing causal constraints as described in F5 and F6 are as follows:
[0141] H0 and self-supervised learning of factor importance are the key to the second core innovation of this invention;
[0142] H1. For each training sample, the true factor importance is calculated using factor perturbation analysis. The secondary topographic factor and the tertiary anthropogenic factor are kept constant. A small perturbation is applied to the primary climate factor, with the perturbation amplitude set at 10% of the factor's standard deviation. This 10% perturbation amplitude is empirically chosen to observe carbon density changes without exceeding a reasonable range. The carbon density change ΔC1 is calculated, and the same operation is performed on other factors to obtain ΔC2 and ΔC3.
[0143] H2. Calculate the true factor importance vector:
[0144]
[0145] This vector represents the percentage contribution of each layer factor to the change in carbon density;
[0146] H3. Add a factor importance prediction branch to the hierarchical prediction model. This branch is a neural network module that takes the feature representations of the three factors as input and outputs the predicted factor importance vector I. pred Softmax normalization ensures that the sum of the three elements is one;
[0147] H4. Define self-supervised loss to measure the difference between predicted importance and true importance:
[0148]
[0149] This loss enables the model to learn the correct distribution of factor importance;
[0150] H5. Define causal constraint loss, using the hinged loss form:
[0151]
[0152] Where θ is the minimum threshold for the importance of climate factors. The threshold is determined by analyzing the statistical distribution of the true importance of climate factors in the training set, and taking its decimal place. Taking the decimal place means that 10% of the samples are allowed to have climate factor importance below the threshold, in order to accommodate individual anomalies. In this embodiment, the decimal place of the true importance of climate factors in the training set is 0.58, so θ = 0.58 is set.
[0153] H6. Construct the total loss function and jointly optimize carbon density prediction, factor importance prediction, and causal constraints:
[0154]
[0155] in The mean squared error loss is for carbon density prediction. The balance coefficients β1 and β2 control the relative importance of the three losses. By trying different combinations on the validation set, the combination that minimizes the root mean square error (RMSE) of the validation set is selected. For this embodiment, the optimal β1 = 0.5 and β2 = 0.1.
[0156] Furthermore, the specific steps of the multi-stage training described in F7 are as follows:
[0157] I0. Model training is divided into three stages. Progressive training ensures that the model maintains both prediction accuracy and conforms to ecological principles.
[0158] I1. First Stage: Pre-training the carbon density prediction branch, optimizing only the carbon density prediction loss, training for thirty epochs with a learning rate of 0.01. This stage enables the model to learn basic carbon density prediction capabilities;
[0159] I2. Second Stage: A factor importance prediction branch and self-supervised loss are added, and the model is trained for twenty epochs with the learning rate reduced to 0.005. This stage allows the model to simultaneously learn carbon density prediction and factor importance prediction.
[0160] Phase 3: Incorporate causal constraint loss, optimize the complete total loss function, and train for twenty epochs with a learning rate of 0.001. This phase imposes explicit constraints on the importance of climate factors to ensure compliance with ecological principles.
[0161] I4. After training, three layers of environmental factors are extracted for each pixel in the target domain, and carbon density is predicted through a hierarchical model.
[0162] Furthermore, step three, the carbon storage calculation, includes:
[0163] J0. For each cell in the target domain, obtain the land type from the land use classification map and extract three-layer factors from the environmental factor data.
[0164] J1. Predict carbon density using a trained hierarchical model to obtain a carbon density distribution map;
[0165] J2. Calculate the carbon storage of a pixel. For a Landsat image with a spatial resolution of 30 meters, the area of each pixel is 0.09 hectares. The carbon storage of a pixel is equal to the carbon density multiplied by 0.09.
[0166] J3. Calculate the total carbon storage by administrative division, generate a spatial distribution map of carbon storage, and save it in GeoTIFF format;
[0167] J4 outputs various statistical results, including total urban carbon storage, average carbon density and carbon storage ratio of each land use type, and spatial distribution characteristics of carbon storage, providing a scientific basis for urban carbon accounting.
[0168] In summary, this embodiment achieves accurate estimation of urban carbon storage by using a fine-grained source domain screening and weighted alignment method based on geographical similarity for cross-regional land use classification, and a self-supervised hierarchical carbon density prediction method based on causal constraints to estimate carbon density. The core idea of this method is to fully utilize local geographical similarity through fine-grained source domain alignment, and to ensure that the model learns the correct factor hierarchy relationships through self-supervised learning based on causal constraints.
[0169] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A method for estimating urban carbon storage based on fine-grained source domain alignment and causal constraints, characterized in that, Includes the following steps: Step 1: Use fine-grained source domain screening based on geographical environment similarity; Step 2: Perform cross-regional land use classification using a weighted alignment method; Step 3: Estimate carbon density using a self-supervised hierarchical carbon density prediction method based on causal constraints; Step 4: Calculate the spatial distribution of urban carbon storage and output the results.
2. The urban carbon storage estimation method based on fine-grained source domain alignment and causal constraints as described in claim 1, characterized in that, Step one specifically includes: A1. Extract geographic environmental feature vector g from remote sensing images of the source and target domains. p =[DEM p Slope p NDVI p LST p ], where DEM p Slope represents the elevation of a pixel. p For the terrain slope of a pixel, NDVI p The normalized vegetation index is calculated using the following formula: NIR p For near-infrared reflectance, Red p For red light band reflectivity, LST p It refers to the surface temperature; A2. For each source domain S i Clustering based on geographical environmental characteristics divides the source domain into several geographically similar sub-regions. Where M i The number of sub-regions of the i-th source domain is determined adaptively using the contour coefficient; A3. Calculate the target domain T and each source domain subregion. The geographical similarity between them is determined by the maximum mean difference (MMD) distance. To measure the differences in the distribution of geographical environmental characteristics.
3. The urban carbon storage estimation method based on fine-grained source domain alignment and causal constraints as described in claim 1, characterized in that, Step two specifically includes: B1. Calculate the alignment weight of each source domain sub-region based on geographical environment similarity. Subregions with high similarity are assigned high weights, while subregions with low similarity are assigned zero weights. B2. Construct a fine-grained weighted domain alignment network, including a feature extractor, a classifier, and a domain discriminator, and use a weighted adversarial training strategy to align the feature distribution of highly similar sub-regions with the target domain. B3. Train the classification model and generate a land use classification map of the target domain.
4. The urban carbon storage estimation method based on fine-grained source domain alignment and causal constraints as described in claim 2, characterized in that, The specific steps for clustering each source domain based on geographical environmental characteristics, as described in A2, are as follows: C1, regarding the source domain S i The geographic environmental feature vectors of all pixels are Z-score standardized to eliminate the influence of differences in the dimensions of different features; C2. Use the K-means clustering algorithm to cluster the source domain pixels, with a cluster size M. i The silhouette coefficient is determined by calculating the silhouette coefficient under different cluster numbers. The silhouette coefficient is defined as follows: Where a p Let b be the average distance from sample p to other samples of the same class. p Choose the number of clusters that maximizes the average silhouette coefficient, given the average distance from sample p to the nearest outlier sample. C3. Obtain the sub-region partitioning results of the source domain. Pixels within each sub-region share similar geographical characteristics.
5. The urban carbon storage estimation method based on fine-grained source domain alignment and causal constraints as described in claim 3, characterized in that, The specific steps for calculating alignment weights and constructing fine-grained weighted domain alignment networks, as described in B1 and B2, are as follows: D1. Calculate the source domain subregion The distance from the maximum mean difference to the target domain T is determined using a Gaussian kernel function: Where g p and g q γ is the geographic environment feature vector, and γ is the kernel bandwidth parameter, which is determined by randomly sampling sample pairs and calculating the median distance. D2. Calculate sub-region alignment weights based on the maximum mean difference distance, using the softmax function: in Let τ be the distance between the source sub-region and the target region, τ be the temperature parameter, τ be the standard deviation of all distances, and N be the number of source regions. D3. Filter high-similarity sub-regions, calculate the third quartile (Q3) of the distance as a threshold, and apply this to the distance... Sub-regions with excessively large environmental differences are excluded by setting their weights to zero. D4. Normalize the non-zero weights so that the sum of all non-zero weights is one; D5. Construct a domain-aligned network, including the feature extractor F. θ Using ResNet-50 and classifier C φ Output six types of land use probabilities and a domain discriminator. Construct an independent discriminator for each source domain sub-region; D6. Construct a fine-grained weighted domain alignment loss function: in For adaptive weights, For source domain sub-regions The adversarial loss against the target domain T; D7. The total loss function combines classification loss and fine-grained weighted domain alignment loss: in λ is the classification loss for the source domain samples, and λ is the balance coefficient, which is determined through validation set performance optimization.
6. The urban carbon storage estimation method based on fine-grained source domain alignment and causal constraints as described in claim 1, characterized in that, Step three specifically includes: E1. Extract environmental factor data and organize it into three levels according to ecological hierarchy: Level 1 factor is climate factor X1, including annual average temperature, annual precipitation, and annual solar radiation; Level 2 factor is topographic factor X2, including altitude, slope, and aspect; Level 3 factor is human activity factor X3, including population density, nighttime light index, and distance from the city center. E2. Construct a hierarchical carbon density prediction model, using a multiplicative-additive combination: C = f1(X1) × f2(X2) + f3(X3) Where f1(X1) is the basic carbon density prediction model based on primary climate factors, f2(X2) is the prediction model with correction coefficients based on secondary topographic factors, f3(X3) is the prediction model with disturbance terms based on tertiary anthropogenic factors, and C is the final predicted carbon density. E3. Construct a self-supervised learning task for factor importance, generate factor importance labels through perturbation analysis, add an importance prediction branch to the hierarchical prediction model, and enable the model to learn carbon density prediction and factor importance prediction simultaneously. E4. Design a causal constraint loss function to ensure that the importance contribution of climate factors conforms to ecological laws; E5. Employ a joint training strategy to simultaneously optimize the carbon density prediction loss and the factor importance prediction loss, and output the carbon density estimation results.
7. The urban carbon storage estimation method based on fine-grained source domain alignment and causal constraints as described in claim 6, characterized in that, The specific steps for constructing the factor importance self-supervised learning task and designing the causal constraint loss function as described in E3 are as follows: F1. For each training sample, keep other factors constant, apply a small perturbation ΔX1 to the primary climate factor X1, with the perturbation amplitude set at 10% of the standard deviation of the factor, and observe the change in carbon density ΔC1; perform the same operation on the secondary topographic factor and the tertiary anthropogenic factor to obtain the changes ΔC2 and ΔC3. F2. Calculate the true factor importance vector: This vector represents the percentage contribution of each layer factor to the change in carbon density, and the sum of the three elements of the vector is one. F3. Add a factor importance prediction branch to the hierarchical prediction model. This branch receives the feature representations of the three factors and outputs the predicted factor importance vector. Normalization using the softmax function ensures that the sum of the three elements is one. F4. Define self-supervised loss to measure the difference between predicted importance and true importance: F5. Define causal constraint loss to ensure that the importance contribution of primary climate factors conforms to ecological principles: Where θ is the minimum threshold for the importance of climate factors, determined based on the statistical distribution of the true importance of climate factors in the training samples, and taken as its tenths. The importance of predicted climate factors; F6. Construct the total loss function and jointly optimize the carbon density prediction and factor importance prediction: in C represents the mean square error loss for carbon density prediction. pred For the predicted carbon density, C true The actual measured carbon density is represented by β1 and β2, which are equilibrium coefficients determined through performance optimization on the validation set.