Block carbon emission prediction method based on process feedback guidance and multi-scale graph neural network
By using process feedback guidance and multi-scale graph neural networks, the problems of ground object interaction and implicit factors in block-scale carbon emission prediction were solved, achieving higher accuracy and more stable carbon emission prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENYANG JIANZHU UNIVERSITY
- Filing Date
- 2026-04-21
- Publication Date
- 2026-07-14
Smart Images
Figure CN122391874A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of remote sensing image classification and extraction technology, and more specifically, to a method for predicting urban carbon emissions using process feedback guidance and multi-scale graph neural networks. Background Technology
[0002] As the primary carrier of human activities, cities are closely linked to carbon emission processes in their spatial morphology. With the continuous advancement of urban renewal, how to scientifically reshape the spatial morphology of city blocks to reduce carbon emissions has become a pressing technical issue for sustainable urban development. At the city block scale, the spatial layout of carbon sources (mainly man-made buildings) and carbon sinks (mainly vegetation) is considered a key factor influencing energy consumption intensity. Their spatial relationship drives regional carbon emission processes by regulating environmental factors such as microclimate and population density. Although existing technologies have made some progress in macro-scale carbon emission analysis, such as carbon emission estimation, carbon metabolism network analysis, and land use carbon response, their prediction accuracy often falls short of the practical needs of urban low-carbon renewal planning when applied to refined city block scale carbon emission prediction.
[0003] Applying existing analytical methods or predictive models to the study of carbon source and sink layout at the street block scale still presents several challenges. Three key issues stand out: First, existing feature representations struggle to capture deep semantic interactions and multi-scale relationships between features. The spatial layout of street blocks involves not only topological interactions between similar features but also antagonism and collaboration between carbon sources and sinks in areas such as land resource competition and microclimate regulation. Traditional models often rely on shallow statistical features (such as POI density and frequency) and local topological connections, neglecting potential semantic relationships between functional facilities (such as industrial chain agglomeration effects). Furthermore, they frequently treat street blocks as isolated units, lacking consideration of their relative position within the overall urban functional structure, resulting in insufficient generalization ability in complex urban environments. Second, existing predictive models struggle to capture the driving effects of implicit environmental factors resulting from spatial layout. While microclimate and population density are direct carriers of carbon emissions, they are implicit feedbacks generated by feature layout, not directly observable street block attributes. Traditional models typically rely on explicit, static data for feature engineering, making it difficult to accurately model the dynamic regulatory role of environmental processes (such as the heat island effect and tidal flow of people) in the interaction with carbon emissions. This ultimately leads to the loss of key explanatory variables in model predictions. Thirdly, high-precision carbon emission training samples at the street-level are scarce. Due to the lack of a high-density network of measured sensors, existing street-level carbon emission data is often difficult to obtain or is obtained through simple averaging, lacking a scientific downscaling mechanism and thus failing to serve as effective supervisory signals for deep learning models. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of the prior art and to provide a method for predicting carbon emissions in urban areas using process feedback guidance and multi-scale graph neural networks.
[0005] Firstly, a method for predicting neighborhood carbon emissions using process feedback guidance and multi-scale graph neural networks is provided, including:
[0006] S1. Use urban road network data to perform topological segmentation of urban space to construct basic street block unit nodes, and allocate the total urban energy statistics value to each street block node based on proxy indicators to generate carbon emission training labels.
[0007] S2. Extract the visual and functional semantic features of the block units and invert the environmental feedback factors used to characterize microclimate processes and social activity processes.
[0008] S3. Using the environmental feedback factor, edge weights between street nodes are dynamically generated through a learnable graph filter to construct a dynamic street graph structure.
[0009] S4. Establish a mutual information maximization model that includes multi-scale relationships, and maximize the mutual information between the street embedding representation and the city global embedding representation through a multi-scale contrastive learning task;
[0010] S5. Construct a collaborative analysis model for heterogeneous multimodal data, build positive and negative sample pairs for environmental feedback factors of different modalities, and achieve multimodal information collaboration by maximizing the similarity of sample pairs within the same spatial unit and minimizing the similarity of sample pairs between different spatial units.
[0011] S6. Using the fusion features obtained in S5 as input, perform multi-task joint training by combining the numerical prediction error of carbon emissions and the contrastive learning loss to obtain the carbon emission prediction results.
[0012] Preferably, S1 includes:
[0013] S101. Use urban road network data to perform morphological expansion and topological closure operations to generate closed street block unit nodes;
[0014] S102. Based on nighttime light data to characterize socioeconomic intensity and building volume to characterize physical carrying capacity, a two-factor proxy allocation model is constructed to calculate the carbon emission allocation weight of each block unit.
[0015] S103. Calculate the true carbon emission label for each block unit based on the allocated weights and the total urban energy statistics.
[0016] Preferably, S2 includes:
[0017] S201. Extract the statistical features, texture features, and spectral index features of the street remote sensing image to form the image physical feature matrix;
[0018] S202. Establish a multi-scale buffer with the street centroid as the origin, extract the interest point features within the buffer, and introduce a multi-head self-attention mechanism for feature encoding to generate functional semantic embeddings.
[0019] S203. Surface temperature is retrieved using thermal infrared remote sensing data and used as an environmental feedback factor to characterize microclimate processes.
[0020] S204. Utilize nighttime light data and demographic data to invert dynamic population density as an environmental feedback factor characterizing social activity processes.
[0021] Preferably, S3 includes:
[0022] S301. For any two street nodes, calculate their normalized absolute difference on the environmental feedback factor, construct an edge feature vector describing the environmental interaction relationship between nodes; and construct an environmental feedback graph tensor based on the edge feature vectors of all node pairs.
[0023] S302. A multi-feature graph signal aggregation strategy is adopted, and a graph filter that depends on the environmental feedback graph tensor is used to dynamically generate the adjacency weights between nodes and update the node embedding.
[0024] S303, Perform intermediate feature generation and aggregation.
[0025] Preferably, S4 includes:
[0026] S401. At the microscale, construct orientation-sensing nodes and distance-sensing nodes centered on ground features, and use graph convolutional networks to aggregate and obtain the street carbon spatial layout embedding.
[0027] S402. At the macro scale, the feature representations of all blocks are aggregated using an area-weighted algorithm to generate a global city summary vector.
[0028] S403. Construct positive and negative sample pairs by combining the street embedding representation with the city global summary vector, and maximize the mutual information between the two by minimizing the contrastive learning loss function.
[0029] As a preferred embodiment, S5 includes:
[0030] S501. Construct positive sample pairs by combining the surface temperature map structure and dynamic population density map structure within the same block, and construct negative sample pairs by combining the surface temperature map structure and dynamic population density map structure between different blocks.
[0031] S502. Construct a similarity measurement mechanism to evaluate the probability that a land surface temperature map and a dynamic population density map belong to the same spatial unit;
[0032] S503. Construct an auxiliary contrast loss function based on the similarity measurement mechanism to bring positive sample pairs closer and negative sample pairs further apart in the feature space, thereby achieving collaborative constraints on heterogeneous modal data.
[0033] As a preferred embodiment, S6 includes:
[0034] S601. Based on the fusion features obtained in the previous steps, obtain the carbon emission prediction value of the corresponding block;
[0035] S602. Construct a multi-task joint loss function, which includes the regression loss of the main task and the contrastive learning loss of the auxiliary task. Learnable parameters are used to dynamically balance the gradient contributions of each task, and multi-task joint training is performed to optimize the model parameters.
[0036] Secondly, a neighborhood carbon emission prediction system guided by process feedback and multi-scale graph neural networks is provided for performing any of the methods described in the first aspect, including:
[0037] The topology segmentation module is used to perform topological segmentation of urban space using urban road network data to construct basic street block unit nodes, and to allocate the total urban energy statistics to each street block node based on proxy indicators to generate carbon emission training labels.
[0038] The extraction module is used to extract the visual and functional semantic features of the block units and to invert the environmental feedback factors used to characterize microclimate processes and social activity processes.
[0039] The generation module is used to dynamically generate edge weights between street nodes using the environmental feedback factor through a learnable graph filter, thereby constructing a dynamic street graph structure.
[0040] A module is established to build a mutual information maximization model that includes multi-scale relationships. This model maximizes the mutual information between the street embedding representation and the city global embedding representation through a multi-scale contrastive learning task.
[0041] The module is used to build a collaborative analysis model for heterogeneous multimodal data. It constructs positive and negative sample pairs from environmental feedback factors of different modalities. By maximizing the similarity of sample pairs within the same spatial unit and minimizing the similarity of sample pairs between different spatial units, it achieves multimodal information collaboration.
[0042] The acquisition module is used to obtain carbon emission prediction results by taking the fusion features obtained from the construction module as input, combining the carbon emission numerical prediction error and the contrastive learning loss for multi-task joint training.
[0043] Thirdly, a computer storage medium is provided, wherein a computer program is stored therein; when the computer program is run on a computer, the computer causes the computer to perform any of the methods described in the first aspect.
[0044] Fourthly, an electronic device is provided, comprising:
[0045] Memory, used to store computer programs;
[0046] A processor for executing the computer program to implement the method as described in any of the first aspects.
[0047] The beneficial effects of this invention are:
[0048] 1. This invention achieves explicit modeling of latent environmental drivers. Unlike traditional models that only consider static spatial distances, this invention constructs an Edge-Varying GNN by inverting surface temperature and dynamic population density. This mechanism can dynamically capture the nonlinear driving effects of microclimate and human activity on carbon emissions, significantly improving the physical interpretability and prediction accuracy of the model in complex urban environments.
[0049] 2. This invention enhances the semantic depth and multi-scale generalization ability of feature representation. Compared with traditional shallow statistical features, it introduces a Transformer self-attention mechanism to extract deep semantic embeddings of POIs; at the same time, it combines multi-scale mutual information maximization (contrastive learning) to force the model to learn both micro-level land cover layout and macro-level urban function positioning, thereby improving the robustness of prediction results. Attached Figure Description
[0050] Figure 1 This is a technical flowchart provided in the embodiments of the present invention;
[0051] Figure 2 This is a flowchart of the construction process of a dynamic street interaction network based on process feedback guidance provided in an embodiment of the present invention;
[0052] Figure 3 This is a flowchart of the multi-scale mutual information maximization model construction process that takes into account local and global relationships, provided by an embodiment of the present invention.
[0053] Figure 4 This is a thematic map of carbon emission prediction results for the main urban area of Shenyang City based on the UIPF-Net model, provided in an embodiment of the present invention.
[0054] Figure 5 This is a thematic map of carbon emission prediction results for the main urban area of Shenyang City based on random forest, provided in an embodiment of the present invention. Detailed Implementation
[0055] The present invention will be further described below with reference to embodiments. The description of the embodiments below is only for the purpose of helping to understand the present invention. It should be noted that those skilled in the art can make several modifications to the present invention without departing from the principle of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.
[0056] Example 1:
[0057] To address the problems of existing technologies, Embodiment 1 of this application provides a method for predicting urban carbon emissions using process feedback guidance and multi-scale graph neural networks, such as... Figures 1-3 As shown, it includes:
[0058] S1. Constructing multi-source sensing street block units and carbon emission proxy labels: Use urban road network data to perform topological segmentation of urban space and construct basic street block unit nodes; obtain the total urban energy statistics, construct a two-factor proxy allocation model based on nighttime light intensity and building volume, adaptively allocate the total carbon emissions to each street block node, and generate carbon emission training labels.
[0059] S1 includes:
[0060] S101. Use urban road network data to perform morphological expansion and topological closure operations to generate closed street block unit nodes.
[0061] For example, different morphological buffer radii can be assigned to different types of roads according to their road classification, and morphological dilation operations can be performed on the road centerlines to enhance the segmentation effect of arterial roads, secondary arterial roads, and local roads on street boundaries. Preferably, the buffer radius of arterial roads can be set to 20-30 meters, the buffer radius of secondary arterial roads can be set to 10-20 meters, and the buffer radius of local roads can be set to 5-10 meters; in some implementations, the above parameters can also be adaptively adjusted according to the road network density and data accuracy of the study area. Subsequently, the closed polygonal regions enclosed by the roads are obtained through set difference operations, serving as the basic street unit.
[0062] Define city administrative regions as a set The road network is a set of lines. Assigning morphological buffer radii to roads of different grades Constructing a morphological expansion set of road space :
[0063]
[0064] in, Represents the morphological dilation operator. For radius . structural elements.
[0065] Next, a set of street blocks is constructed, and each street block is abstracted as a node in a graph structure.
[0066] For example, each enclosed street block is treated as a basic spatial analysis unit, and a unique number, geometric boundary, area attribute, and centroid coordinates are assigned to it, thus forming a street block-scale node database. This provides a spatial carrier for subsequent feature extraction, graph relationship construction, and carbon emission prediction. Furthermore, abnormal street block units that are too small, have overly fragmented shapes, or lack key attributes can be removed or merged to improve the stability of the street block map construction.
[0067] Generate the basic street unit set V using set difference operations:
[0068] in, The first one obtained through topological closure operation A closed polygonal street block unit.
[0069] S102. Based on nighttime light data (NTL) to characterize socioeconomic intensity and building volume (Vol) to characterize physical carrying capacity, a two-factor proxy allocation model is constructed to calculate the carbon emission allocation weight of each block unit.
[0070] Specifically, considering the difficulty in directly obtaining accurate carbon emission observations at the street-level, this embodiment uses nighttime light intensity to represent socioeconomic activity intensity and building volume to represent the street's physical carrying capacity, coupling the two to construct a two-factor proxy allocation model. Based on nighttime light data (NTL) representing socioeconomic intensity and building volume (Vol) representing physical carrying capacity, the model calculates the... Block Units Allocation weights The formula is as follows:
[0071]
[0072] in, The total number of blocks, As a regulating factor; preferably, A value between 0.4 and 0.6 can be used; in this specific embodiment, A value of 0.5 is appropriate to take into account the combined impact of socio-economic activity intensity and building development intensity on the carbon emissions of the neighborhood.
[0073] Based on the total urban energy statistics Calculate the true carbon emission label for each neighborhood. A labeled spatial database of carbon emissions from neighborhoods will be constructed. Furthermore, building volume can be estimated from the building footprint area and building height; when building height data is difficult to obtain directly, the number of floors, floor area ratio, or other indicators that can reflect construction intensity can be used for approximate characterization.
[0074] S103. Calculate the true carbon emission label for each block unit based on the allocated weights and the total urban energy statistics.
[0075] Specifically, based on the total urban energy statistics ; Calculate the true carbon emission label for each neighborhood
[0076] Construct a labeled spatial database of neighborhood carbon emissions.
[0077] S2. Multidimensional feature extraction and environmental process parameter inversion: Extract visual texture features and functional semantic features of points of interest (POIs) from street remote sensing images; use multi-source remote sensing data to invert surface temperature and dynamic population density as environmental feedback factors characterizing microclimate processes and social activity processes.
[0078] S2 includes:
[0079] S201. Extract the statistical features, texture features, and spectral index features of the street remote sensing image to form the image physical feature matrix.
[0080] Specifically, mean layer, standard deviation, skewness, and gray-level co-occurrence matrix (GLCM) texture features can be extracted from the remote sensing images corresponding to each block to characterize the surface cover structure, texture complexity, and spatial heterogeneity within the block. Simultaneously, the normalized difference in vegetation index (NDVI) and normalized difference in building index (NDBI) are calculated to characterize the physical properties of carbon sink and carbon source features within the block, respectively. Furthermore, the GLCM texture features may include at least one of contrast, energy, entropy, homogeneity, and correlation; the specific statistical window can be set according to the image spatial resolution.
[0081] Furthermore, to quantify the physical properties of carbon sources and sinks within the block, the Normalized Difference Vegetation Index (NDVI) and Normalized Difference Building Index (NDBI) were calculated using the following formulas:
[0082]
[0083] in, It is in the near-infrared band. It is in the red light band. The reflectance value is the shortwave infrared band; the spectral index is concatenated with the texture features to form the image physical feature matrix. .
[0084] S202. Establish a multi-scale buffer with the street centroid as the origin, extract the interest point features within the buffer, and introduce a multi-head self-attention mechanism for feature encoding to generate functional semantic embedding.
[0085] For example, a multi-scale buffer set is established with the street centroid as the origin. Density and diversity characteristics of POIs are statistically analyzed within each buffer scale to characterize the functional attributes of the street in terms of commerce, residence, industry, and public services. Preferably, the radius of the multi-scale buffer can be set to 100 meters, 300 meters, and 500 meters, or other buffer distances adapted to the average scale of the street. Subsequently, an attention-based feature encoder is introduced to map the initial POI statistical features into functional embeddings containing contextual semantic relationships, thereby enhancing the street's functional expressiveness.
[0086] Specifically, S202 includes:
[0087] S2021, Construction of multi-scale buffer feature matrix.
[0088] For example, using the centroid of the block as the origin, establish A set of buffers with different radii (e.g., 100m, 300m, and 500m); for the first The block in the The first in the buffer Similar to POIs, calculate density and diversity metrics to construct initial feature vectors. The formula is as follows:
[0089]
[0090]
[0091] in, For the number of POIs, Where is the radius of the buffer zone. The diversity is defined using Shannon Entropy to represent the proportion of this type of POI within the buffer.
[0092] S2022, Semantic Embedding of POIs in Urban Environments.
[0093] Specifically, a multi-head self-attention mechanism is introduced as a feature encoder to map the initial features into semantic embeddings containing contextual information; firstly, the feature vectors are... Linear projection of the query vector Key vector Sum value vector :
[0094] S2023, Semantic Feature Aggregation.
[0095] Specifically, calculate the first The weighted output of each attention head, and all The outputs of each element are concatenated to obtain the final urban environment semantic embedding. The formula is as follows:
[0096]
[0097]
[0098] in, For feature dimension, For the number of attention heads, This represents a vector concatenation operation. It is a normalized exponential function.
[0099] S203. Surface temperature is retrieved using thermal infrared remote sensing data and used as an environmental feedback factor to characterize microclimate processes.
[0100] Specifically, a single-window algorithm is used to retrieve land surface temperature using the thermal infrared band. First, the digitally quantized values of the thermal infrared band are converted into spectral radiance. Then, the radiance temperature is calculated using the Planck function. Finally, correction is performed using the surface emissivity to obtain the corresponding land surface temperature value for each block. The formula is as follows:
[0101]
[0102] in, The center wavelength, These are physical constants. The surface emissivity is used; cloud, shadow and outlier removal can be performed on the inversion results, and the mean or area-weighted average value within the block area can be used as the surface temperature characterization value of the block.
[0103] S204. Utilize nighttime light data and demographic data to invert dynamic population density as an environmental feedback factor characterizing social activity processes.
[0104] For example, nighttime light data is used as a proxy variable for human activity intensity, and combined with road network density and land use data to construct a spatial population model; then, population statistics at the administrative division level are downscaled and allocated to street block units to obtain the dynamic population density of each block. The formula is as follows:
[0105] in, The average light intensity of the block, The area of the block; the inverted area and Combining these elements to form an environmental feedback feature vector This is used for generating the edge weights of the subsequent graph network.
[0106] Furthermore, the surface temperature and dynamic population density obtained from the inversion are combined to form a street environment feedback feature vector, which is used for the construction of subsequent street interaction edge features and the generation of dynamic adjacency relationships.
[0107] S3. Construct a process feedback-guided street interaction network: Construct a relational change graph neural network, utilize the differences in environmental feedback factors obtained from step 2, and dynamically generate edge weights between street nodes through a learnable graph filter to construct a dynamic street graph structure that takes into account the implicit environmental driving effects.
[0108] S3 further includes: using the street block units obtained in S1 as nodes, and concatenating the visual texture features, physical attribute features, functional semantic features, and environmental feedback features extracted in S2 to form the initial feature representation of the street block node. Further, the POI map features and image map features can be encoded separately and then fused at the street block scale to obtain a multimodal street block representation. Preferably, the hidden feature dimension can be set to 64; in other embodiments, it can also be set to 32, 128, or other reasonable values depending on the data scale and hardware conditions.
[0109] S3 specifically includes:
[0110] S301. Define a filter function based on environmental differences; for any two block nodes, calculate their normalized absolute difference on the environmental feedback factor, construct an edge feature vector describing the environmental interaction relationship between nodes; and construct an environmental feedback graph tensor based on the edge feature vectors of all node pairs.
[0111] Specifically, for any two street nodes and Calculate its surface temperature ( ) ) and dynamic population density ( The normalized differences on the graph are calculated, and the above difference indicators are concatenated to form the environmental interaction edge feature vector between node pairs. This edge feature is used to reflect the feedback differences of different neighborhoods in microclimate processes and social activity processes. The formula is as follows:
[0112]
[0113]
[0114] By concatenating the above difference indicators, an edge feature vector describing the environmental interaction relationship between nodes is constructed. :
[0115]
[0116] Construct an environment feedback graph tensor based on the edge feature vectors of all node pairs. (in (where 2 represents the number of nodes and 2 represents the environmental factor dimension), this tensor This will be used as input to subsequent graph filters to generate dynamic adjacency weights driven by environmental gradients.
[0117] S302. A multi-feature graph signal aggregation strategy is adopted, and a graph filter that depends on the environmental feedback graph tensor is used to dynamically generate the adjacency weights between nodes and update the node embedding.
[0118] In S302, a learnable graph filter is introduced to dynamically adjust the edge connection strength according to the environmental differences between nodes, so that the street graph structure is no longer limited to fixed topological adjacency relationships, but can explicitly reflect the potential interaction driven by environmental feedback.
[0119] Specifically, a multi-feature map signal aggregation strategy is adopted, in which different types of features are processed by independent graph filters, and a trainable weighting matrix is used. Fusion, the first Layer node embedding The calculation formula is:
[0120]
[0121]
[0122] in, It is a non-linear activation function. Let the order be the filter order. Indicates the first An information aggregation operator with different hop counts within a layer. This operator is a common multi-hop neighborhood aggregation function in graph neural networks. Representation matrix of The power of n, in its physical sense, is the result of performing matrix multiplication on the nth node. The aggregation and transmission of information in the spatial neighborhood; particularly, when hour, , The identity matrix is of the same dimension, representing the preservation of the local input features of each block node. A two-layer graph neural network can be used to perform node relationship propagation, with each layer followed by a non-linear activation function; in one implementation, the block graph can be undirected to enhance the consistency of spatial interaction representation.
[0123] S303, Perform intermediate feature generation and aggregation.
[0124] Specifically, the formula for this step is as follows:
[0125]
[0126]
[0127] in, Indicates the first Output features in the layer Input features Intermediate eigenvalues between Indicates from input features To output features The parameters of the graph filter.
[0128] After executing S3, the street-level interaction representation learning is completed. Through multi-layer graph signal propagation and nonlinear mapping, neighborhood node information is gradually aggregated, so that each street node not only contains local feature layout information, but also incorporates neighborhood interaction relationships modulated by environmental process feedback, thereby improving the ability to characterize complex carbon emission mechanisms. Furthermore, during the node update process, normalization constraints, feature scale alignment, or residual fusion strategies can be combined to reduce the oversmoothing phenomenon in graph neural networks.
[0129] S4. A regional feature aggregation method that considers multi-scale spatial dependence establishes a mutual information maximization model for the multi-scale relationship between "landform-block-city". Micro-level landform layout features are aggregated within blocks, while a multi-scale contrastive learning task is constructed to maximize the mutual information between the block embedding representation and the city's global summary.
[0130] S4 includes:
[0131] S401. At the microscale, construct direction-aware nodes and distance-aware nodes centered on ground features, and use graph convolutional networks to aggregate and obtain the carbon spatial layout embedding of the block.
[0132] Specifically, at the microscale, modeling the layout of features within a block allows for the construction of orientation-aware nodes and distance-aware nodes centered on features. These nodes characterize the spatial relative positions, distribution patterns, and potential mechanisms of carbon source and sink features. Carbon source features can include buildings, roads, and other heavily constructed surfaces, while carbon sink features can include vegetation, water bodies, and other ecological spatial elements. Orientation-awareness describes the locational relationships between features, while distance-awareness describes their proximity and dispersion. Furthermore, a graph convolutional network is used to aggregate the microscopic feature relationships within a block, resulting in a block layout embedding representation that characterizes the spatial configuration of carbon sources within the block. Graph-level pooling is performed on POI maps and image maps to obtain multimodal embedding results for the corresponding blocks, which are then jointly modeled with the block relationship graph representation.
[0133] For example, construct direction-aware nodes and distance-aware nodes centered on ground features, based on the ground feature nodes. Based on its spatial location, its neighboring nodes are divided into different directional sector sets. and different distance annular sets Then, the directional embedding vectors are calculated separately. and distance embedding vector The node embeddings of the central node are merged and updated using a graph convolutional network. The specific formula is as follows:
[0134]
[0135]
[0136]
[0137] in, Represents a non-linear activation function. For index number, Indicates the central node The set of neighboring nodes within a specific directional partition. Indicates the central node The set of neighboring nodes within a specific distance ring; This represents the total number of neighbors of a node; Neighboring nodes Input features, For trainingable weight matrices extracted for directional features, This is a trainable weight matrix for distance feature extraction. and For the feature fusion weight matrix, The weights are trained; then, a graph convolutional network is used to aggregate the spatial layout embedding of carbon sources in the neighborhood.
[0138] S402. At the macro scale, the feature representations of all blocks are aggregated using an area-weighted algorithm to generate a global city summary vector.
[0139] Specifically, at a macro scale, the representations of all street nodes are aggregated using area weighting to generate a global city summary vector. This is used to characterize the overall functional structure, spatial organization, and carbon metabolism pattern of a city. The formula is as follows:
[0140]
[0141] in, Embedded into the neighborhood, It is the first The percentage of the total area of each block in the city. It is the Sigmoid activation function.
[0142] S403. Construct positive and negative sample pairs by combining the street embedding representation with the city global summary vector, and maximize the mutual information between the two by minimizing the contrastive learning loss function.
[0143] Specifically, positive sample pairs are constructed by combining the street-level embedding representation with the city-wide summary vector, and negative sample pairs are constructed by perturbation, shuffling, or replacement. The contrastive learning loss function is used to maximize the mutual information between the local street representation and the city-wide representation, thereby enhancing the model's ability to learn the "relative location of the street in the overall city system".
[0144] S5. Construct a heterogeneous multimodal data collaborative analysis model, build positive and negative sample pairs for environmental feedback factors of different modalities, and achieve multimodal information collaboration by maximizing the similarity of sample pairs within the same spatial unit and minimizing the similarity of sample pairs between different spatial units.
[0145] S5 includes:
[0146] S501. Construct heterogeneous modal sample pairs: construct positive sample pairs by combining the surface temperature map structure and the dynamic population density map structure within the same block, and construct negative sample pairs by combining the surface temperature map structure and the dynamic population density map structure between different blocks.
[0147] Specifically, both the surface temperature map structure and the dynamic population density map structure use the same basic street block unit as the node set. The surface temperature map structure uses the normalized absolute difference in surface temperature between nodes. The edge weights are used to characterize the spatial interactions of the microclimate thermal environment; the dynamic population density map structure uses the normalized absolute difference in dynamic population density between nodes. As edge weights, they are used to construct spatial interaction relationships that characterize the intensity of social activities.
[0148] In the feature extraction stage, the multimodal features of the blocks obtained in the previous step are used as the initial input to the nodes. Two graph neural network branches with independent parameters are used for processing. The edge weights of two different graph structures are used to perform multi-level graph signal propagation, aggregation, and graph-level pooling operations to output the corresponding blocks. Surface temperature map embedding vector and dynamic population density map embedding vector Subsequently, units belonging to the same neighborhood will be... of and Construct positive sample pairs, belonging to different block units (such as blocks) With the neighborhood )of and Construct negative sample pairs.
[0149] S502. Construct a similarity measurement mechanism: Evaluate the probability that the surface temperature map and the dynamic population density map belong to the same urban functional area through the similarity matrix, and learn the consistent expression and difference boundary between heterogeneous modal data by maximizing the similarity score of positive sample pairs and minimizing the similarity score of negative sample pairs.
[0150] For example, during training, a similarity matrix is constructed. To evaluate the probability that a surface temperature map and a dynamic population density map belong to the same urban neighborhood, and by maximizing... The similarity scores of each positive sample pair are minimized. The objective function is optimized using the similarity scores of negative sample pairs. The formula for calculating the similarity score is as follows:
[0151]
[0152] in, For the first Embedding vectors of surface temperature maps for each block For the first Embedding vectors of dynamic population density maps within a block These are trainable temperature parameters.
[0153] S503. Constructing an auxiliary contrastive loss function: Based on the similarity score matrix, an auxiliary contrastive loss function is constructed to make the surface temperature embedding and dynamic population density embedding within the same block more similar in the feature space, while maintaining high discriminative power for heterogeneous modal representations between different blocks, thereby achieving information interaction and collaborative constraints under spatial semantic injection. The formula is as follows:
[0154]
[0155] in, For the number of blocks, This is a similarity measurement function.
[0156] S6. Using the fusion features obtained in S5 as input, perform multi-task joint training by combining the numerical prediction error of carbon emissions and the contrastive learning loss to obtain the carbon emission prediction results.
[0157] S6 includes:
[0158] S601. Based on the fusion features obtained in the previous steps, obtain the carbon emission prediction value of the corresponding block.
[0159] Specifically, a carbon emission regression prediction module is constructed. The fused features obtained in the preceding steps are input into this module, which outputs the predicted carbon emission values for the corresponding street blocks. This module consists of two feature mapping layers and one output layer. The hidden layer feature dimension is set to 64, and the output layer dimension is set to 1, used to obtain continuous predicted carbon emission values for each street block unit. To enhance the model's training stability, a non-linear activation function is introduced between the hidden layers, and normalization processing is combined to improve the representational ability of the fused features from different modalities.
[0160] S602. Construct a multi-task joint loss function, which includes the regression loss of the main task and the contrastive learning loss of the auxiliary task. Learnable parameters are used to dynamically balance the gradient contributions of each task, and multi-task joint training is performed to optimize the model parameters.
[0161] For example, the prediction error of carbon emissions in the urban area is used as the primary task loss, and the multi-scale mutual information comparison loss in S5 is used as the auxiliary task loss. A joint optimization approach is adopted to update the model parameters synchronously, so as to mitigate the problem that a single supervision signal is insufficient to express the complex carbon emission mechanism of urban areas. An automatic weighted loss strategy is used to dynamically balance the primary and auxiliary tasks, and the total loss function of the multiple tasks is... As shown in the following formula:
[0162]
[0163] in, The mean squared error regression loss for the main task. Comparative learning loss for auxiliary tasks; and It is a learnable variance parameter during model training, used to dynamically balance the gradient contributions of the main task and auxiliary tasks, and to prevent negative transfer in multi-task training.
[0164] Finally, the model training and prediction output are completed.
[0165] Specifically, based on the constructed spatial database of neighborhood carbon emissions, the model is trained end-to-end, ultimately outputting refined carbon emission prediction results for each neighborhood unit within the target area. During model training, the optimizer is set to Adam, the learning rate to 0.001, the batch size to 16, and the training epochs to 200. To prevent overfitting in the later stages of training, the optimal model parameters are saved during training by incorporating changes in the validation set loss.
[0166] Furthermore, the samples are divided into training and testing sets, with a ratio of 8:2. After training, the mean absolute error, root mean square error, and coefficient of determination are used to evaluate the results. As a performance evaluation metric for the model, its expressions are as follows:
[0167]
[0168]
[0169]
[0170] in, Indicates the number of test samples. This represents the average of actual carbon emissions.
[0171] By combining numerical supervision of carbon emissions with multi-scale structural constraints, S6 enables the model to maintain the accuracy of carbon emission prediction at the street level while further enhancing its comprehensive representation of local spatial relationships, neighborhood interaction processes, and overall urban structural characteristics, thereby improving the reliability and stability of street-scale carbon emission prediction results.
[0172] Example 2:
[0173] Building upon Example 1, Example 2 of this application provides a more specific method for predicting neighborhood carbon emissions using process feedback guidance and multi-scale graph neural networks, including:
[0174] S1. Acquire multi-source basic data for the target area and construct street block units and street block carbon emission training labels. Acquire road network data, nighttime light data, building outline data, building height data, energy statistics data, remote sensing image data, point of interest (POI) data, land use data, and population statistics at the administrative division scale for the target area; perform topological segmentation of the target area based on the road network to form street block units; further, combine nighttime light intensity and building volume to construct a two-factor proxy allocation model, downscale and allocate the total energy statistics of the target area to each street block unit, and generate street block-scale carbon emission training labels.
[0175] Specifically, road network data is processed in a hierarchical manner according to arterial roads, secondary arterial roads, and local roads. The morphological buffer radius of arterial roads is set to 25 meters, that of secondary arterial roads to 15 meters, and that of local roads to 8 meters. Closed block boundaries are formed by performing morphological dilation and topological closure on roads of different levels. Then, a set of basic block units for the target area is generated through set difference operations. Abnormal block units with excessively small areas, fragmented boundaries, or missing key attributes are removed or merged to improve the stability of subsequent modeling. Furthermore, to address the difficulty in obtaining real observation samples of carbon emissions at the block scale, a two-factor proxy allocation model of nighttime light intensity and building volume is constructed. Nighttime light intensity represents the intensity of socio-economic activities, and building volume represents the physical carrying capacity of the block; the building volume is obtained by multiplying the building base area and building height within the block and then summing the results. When some building height data is missing, equivalent approximations are made using the number of floors or development intensity indicators. Adjustment factors... The value was set to 0.5 to account for the combined impact of nighttime light intensity and building volume on street block carbon emissions. The weighting of each street block was then calculated, and combined with the total energy statistics for the target area, street block-scale carbon emission training labels were generated. To mitigate the influence of extreme values, the street block carbon emission labels underwent a square root transformation before training.
[0176] In this implementation, the carbon emission labels of the neighborhoods are first subjected to a square root transformation to reduce the impact of extreme values on the training process. Furthermore, outliers are pruned according to the 0.999 quantile to improve the stability of the supervision signal. The corrected neighborhood carbon emission labels are used as the supervision targets for subsequent model training.
[0177] S2. Extract multimodal features of the block and invert environmental process parameters to construct block environmental feedback features. Based on the block units generated in S1, extract the visual texture features and POI functional semantic features from the block remote sensing imagery, and combine thermal infrared remote sensing information and a population spatialization model to invert the block surface temperature and dynamic population density, thereby constructing an environmental feedback feature vector. Specifically, the implementation process of block multimodal feature extraction and environmental process parameter inversion is as follows:
[0178] S201: Extract statistical and textural features from street remote sensing images.
[0179] Using the street blocks obtained from S1 as the basic spatial analysis object, the remote sensing image corresponding to each street block is cropped, and the mean brightness, standard deviation, skewness, and gray-level co-occurrence matrix (GLCM) texture features are extracted. The GLCM texture features include contrast, energy, entropy, homogeneity, and correlation. To balance texture representation capability and computational stability, the texture window size is set to... Simultaneously, the Normalized Difference Vegetation Index (NDVI) and Normalized Difference Building Index (NDBI) were calculated based on multispectral bands to characterize the features of carbon sink and carbon source features within the block, respectively. Finally, the statistical features, texture features, NDVI features, and NDBI features were stitched together to form the initial feature vector of the block image.
[0180] S202: Extract the functional semantic features of the street blocks.
[0181] A multi-scale buffer set is constructed with the street centroid as the origin, and the buffer radii are set to 100 meters, 300 meters, and 500 meters, respectively. Within each buffer, the category density and diversity indices of POIs are statistically analyzed to characterize the functional attributes of the street, such as commercial, residential, industrial, and public service functions. Further, the statistical features of POIs at different scales are input into a multi-head self-attention encoder for feature mapping to obtain functional semantic embeddings that include contextual relationships. In this embodiment, the number of self-attention heads is set to 2, and the hidden feature dimension is set to 64 to ensure a balance between semantic feature extraction capability and model complexity. Finally, the semantic embeddings of POIs at different scales are fused to obtain the functional semantic feature representation of the street.
[0182] S203: Inverted street surface temperature.
[0183] Based on thermal infrared remote sensing data of the target area, a single-window algorithm is used to retrieve the surface temperature of a block. First, the digital quantization values of the thermal infrared bands are converted into spectral radiance. Second, the radiance temperature is calculated using the Planck function. Finally, corrections are made based on surface emissivity to obtain the surface temperature result at the block scale. In this embodiment, the original surface temperature retrieval results are processed to remove cloud cover, shadows, and outliers, and the area-weighted average value within the block area is used as the surface temperature characterization value for that block, thus forming a microclimate process feedback feature.
[0184] S204: Inverting the dynamic population density of the neighborhood.
[0185] This study utilizes nighttime light data as a proxy variable for human activity intensity and constructs a spatial population model by combining street network density and land use information. Population statistics at the administrative division level are downscaled and allocated to street units to obtain the dynamic population density of each street. In this implementation, the product of the average nighttime light intensity and the street network density is used to represent the level of human activity, and then normalized by combining this with the street area to form a dynamic population density index for each street. To ensure comparability between different streets, the dynamic population density results are standardized.
[0186] S205: Constructing the street environment feedback feature vector.
[0187] The street surface temperature obtained from S203 is combined with the street dynamic population density obtained from S204 to form a street environment feedback feature vector.
[0188] S3. Construct a dynamic street interaction network guided by process feedback based on environmental feedback differences.
[0189] Based on the street image features, POI semantic features, and environmental feedback features obtained from S2, a street interaction map is constructed, and a relational change graph neural network is used to model the environmentally driven relationships between streets. Specifically, each street is first treated as a node, and initial connections are established between spatially adjacent streets or streets with close centroid distances. Then, the normalized absolute differences between street nodes in terms of surface temperature and dynamic population density are calculated, and these are concatenated to form an edge feature vector, which is used to describe the interaction differences between streets caused by environmental gradients.
[0190] In this embodiment, the street interaction graph is constructed as an undirected graph to enhance the consistency of neighborhood spatial interaction representation. The street interaction network uses a two-layer relational change graph neural network for feature propagation, with the hidden feature dimension set to 64 and the graph filter order set to 2. Through a learnable graph filter, the connection strength between nodes is dynamically adjusted according to edge features, so that street relationships are no longer limited to static spatial adjacency, but can reflect the potential street interaction mechanisms driven by both microclimate processes and social activity processes. Furthermore, a normalization and residual fusion strategy is introduced during the node update process to reduce the over-smoothing phenomenon in the graph neural network and improve the stability of street interaction representation learning.
[0191] S4. A regional correlation feature aggregation method that takes into account multi-scale spatial dependence is used to establish a mutual information maximization model for the multi-scale relationship between "land features, blocks, and cities".
[0192] At the microscale, the spatial relationships between carbon source and sink features within a block are modeled. At the macroscale, weighted aggregation of all block node representations yields a global city summary vector, and contrastive learning is used to enhance the consistency between local block representations and the overall city representation. Specifically, firstly, direction-aware nodes and distance-aware nodes centered on features are constructed within the block. Carbon source features include buildings, roads, and other high-intensity construction surfaces, while carbon sink features include vegetation, water bodies, and other ecological spatial elements. Local micro-feature relationships are aggregated using a graph convolutional network to obtain a block-level carbon spatial layout embedding. Furthermore, graph-level pooling is performed on the POI graph representation and image graph representation respectively, and they are fused with the block relationship features output from the block interaction graph to form a comprehensive block embedding representation.
[0193] Before modal fusion, L2 normalization was performed on both the POI image embedding and the image image embedding to reduce the impact of scale differences between different modal features on the fusion result. A numerical stabilization term was added during the normalization process. To avoid division by zero, the normalized features of the two modalities are then input into the cross-attention fusion module for alignment and interaction.
[0194] In this implementation, the number of attention heads in the cross-attention module is set to 2, and the dimension of the fused hidden features is set to 64. At the city scale, an area-weighted aggregation method is used to generate a global city summary vector.
[0195] Furthermore, positive sample pairs are constructed by combining the street block embedding representation with the city global summary vector, and negative sample pairs are constructed by perturbation, shuffling, or replacement. Multi-scale contrastive loss is used to maximize the mutual information between the local street block representation and the city global representation, thereby enhancing the model's ability to learn the functional positioning of streets in the overall urban carbon metabolism system.
[0196] S5. Collaborative analysis of heterogeneous multimodal data based on information interaction under spatial semantic injection.
[0197] A collaborative mechanism for the interaction of land surface temperature and dynamic population density is constructed. During training, a similarity matrix is built. To evaluate the probability that surface temperature maps and dynamic population density maps belong to the same urban functional area, and to maximize The similarity scores of each positive sample pair are minimized. The similarity scores of negative sample pairs are used to optimize the objective function, thereby enhancing the model's ability to characterize the synergistic relationship between neighborhood microclimate processes and social activity processes.
[0198] Specifically, positive sample pairs are constructed by combining surface temperature map structures and dynamic population density map structures within the same neighborhood, while negative sample pairs are constructed by combining surface temperature map structures and dynamic population density map structures between different neighborhoods. This guides the model to learn the consistent representation and difference boundaries between heterogeneous modal data. Through the aforementioned similarity measure, the degree of matching between two types of heterogeneous modalities within the same neighborhood in the semantic space can be quantitatively represented.
[0199] In this embodiment, an auxiliary contrastive loss function is further constructed based on the similarity score matrix. This makes the surface temperature embedding and dynamic population density embedding within the same block more similar in the feature space, while maintaining high discriminativeness of heterogeneous modal representations between different blocks. This achieves information interaction and collaborative constraints under spatial semantic injection. Through this step, the model's ability to express the coupling relationship of environmental processes can be further enhanced, providing more discriminative heterogeneous multimodal fusion features for subsequent block carbon emission regression prediction.
[0200] S6. Complete the prediction of carbon emissions in the block based on multi-task joint constraints.
[0201] The fused features obtained from S5 are input into the carbon emission regression prediction module to perform regression prediction on the carbon emission values of the block. Simultaneously, a multi-scale mutual information comparison loss is used for joint optimization to achieve block-scale carbon emission prediction. Specifically, the carbon emission regression prediction module consists of two feature mapping layers and one output layer. The hidden layer feature dimension is set to 64, and the output layer dimension is set to 1, obtaining continuous carbon emission prediction values for each block unit. The main task loss uses mean squared error regression loss; the auxiliary task loss uses multi-scale mutual information comparison loss, with an auxiliary task weight set to 0.01. In this embodiment, the model training uses the Adam optimizer, with a learning rate of 0.001, a batch size of 16, and 200 training epochs. The constructed block samples are divided into training and test sets, with a ratio of 8:2. During training, the optimal model parameters are saved by considering the changes in test set loss. After training, the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination are used to calculate the optimal model parameters. The predictive performance of the model was evaluated. Furthermore, the trained model was used to infer carbon emissions from all blocks in the target area, yielding refined spatial distribution results of carbon emissions at the block scale. The results show that this method can effectively improve the stability and accuracy of block carbon emission prediction, and can simultaneously characterize the local feature layout relationships, neighborhood environmental interactions, and overall urban functional structure characteristics, providing reliable data support for urban low-carbon renewal planning and spatial optimization decisions.
[0202] As shown in Table 1, in this embodiment, the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) are used. 2 The model's predictive performance was evaluated, with MAE and RMSE expressed in tons of CO2. To verify the effectiveness of the proposed method, a Random Forest (RF) regression model was set as a baseline, and training and testing were performed using the same training and test set partitioning as the proposed model. After training, the model's results on the test set were: Mean Absolute Error (MAE) of 369.2881 tons of CO2, Root Mean Square Error (RMSE) of 513.1487 tons of CO2, and Coefficient of Determination (R²). 2 The mean absolute error (MAE) of the random forest baseline model on the test set was 0.8250. The results showed that the MAE was 484.0825 tons of CO2, the root mean square error (RMSE) was 680.7219 tons of CO2, and the coefficient of determination (R²) was 0.6920. In comparison, the proposed model reduced CO2 by 114.7944 tons of MAE and 167.5732 tons of RMSE compared to the random forest baseline, and the R² was 0.6920. 2The improvement of 0.1330 in the indicator indicates that the proposed method can more fully explore the local feature layout relationships, neighborhood environmental interactions, and overall urban functional structure characteristics of blocks, thereby achieving higher accuracy and stronger stability in block-scale carbon emission prediction. Furthermore, the trained model is used to infer carbon emissions from all blocks in the target area, yielding refined spatial distribution results of carbon emissions at the block scale. Analysis shows that high-carbon emission blocks are mainly distributed in areas with high construction intensity, dense population activity, and significant thermal environment, while low-carbon emission blocks are mainly distributed in areas with high vegetation cover, proximity to water bodies, or lower development intensity. These results are in good agreement with the urban spatial functional distribution and surface environmental characteristics, indicating that the proposed method can effectively characterize the spatial differences in block carbon emissions and has good spatial interpretability and predictive reliability.
[0203] Table 1. Comparison of carbon emission prediction results of different methods on the Shenyang urban dataset.
[0204]
[0205] It should be noted that the parts in this embodiment that are the same as or similar to those in Embodiment 1 can be referred to each other, and will not be repeated in this application.
[0206] Example 3:
[0207] Based on Example 2, Example 3 of this application provides a street block carbon emission prediction system guided by process feedback and multi-scale graph neural network, including:
[0208] The topology segmentation module is used to perform topological segmentation of urban space using urban road network data to construct basic street block unit nodes, and to allocate the total urban energy statistics to each street block node based on proxy indicators to generate carbon emission training labels.
[0209] The extraction module is used to extract the visual and functional semantic features of the block units and to invert the environmental feedback factors used to characterize microclimate processes and social activity processes.
[0210] The generation module is used to dynamically generate edge weights between street nodes using the environmental feedback factor through a learnable graph filter, thereby constructing a dynamic street graph structure.
[0211] A module is established to build a mutual information maximization model that includes multi-scale relationships. This model maximizes the mutual information between the street embedding representation and the city global embedding representation through a multi-scale contrastive learning task.
[0212] The module is used to build a collaborative analysis model for heterogeneous multimodal data. It constructs positive and negative sample pairs from environmental feedback factors of different modalities. By maximizing the similarity of sample pairs within the same spatial unit and minimizing the similarity of sample pairs between different spatial units, it achieves multimodal information collaboration.
[0213] The acquisition module is used to obtain carbon emission prediction results by taking the fusion features obtained from the construction module as input, combining the carbon emission numerical prediction error and the contrastive learning loss for multi-task joint training.
[0214] It should be noted that the system provided in this embodiment is the system corresponding to the method provided in embodiment 2. Therefore, the parts in this embodiment that are the same as or similar to those in embodiment 2 can be referred to each other, and will not be described again in this application.
Claims
1. A method for predicting urban carbon emissions using process feedback guidance and multi-scale graphical neural networks, characterized in that, include: S1. Use urban road network data to perform topological segmentation of urban space to construct basic street block unit nodes, and allocate the total urban energy statistics value to each street block node based on proxy indicators to generate carbon emission training labels. S2. Extract the visual and functional semantic features of the block units and invert the environmental feedback factors used to characterize microclimate processes and social activity processes. S3. Using the environmental feedback factor, edge weights between street nodes are dynamically generated through a learnable graph filter to construct a dynamic street graph structure. S4. Establish a mutual information maximization model that includes multi-scale relationships, and maximize the mutual information between the street embedding representation and the city global embedding representation through a multi-scale contrastive learning task; S5. Construct a collaborative analysis model for heterogeneous multimodal data, build positive and negative sample pairs for environmental feedback factors of different modalities, and achieve multimodal information collaboration by maximizing the similarity of sample pairs within the same spatial unit and minimizing the similarity of sample pairs between different spatial units. S6. Using the fusion features obtained in S5 as input, perform multi-task joint training by combining the numerical prediction error of carbon emissions and the contrastive learning loss to obtain the carbon emission prediction results.
2. The method for predicting urban carbon emissions using process feedback guidance and multi-scale graph neural networks according to claim 1, characterized in that, S1 includes: S101. Use urban road network data to perform morphological expansion and topological closure operations to generate closed street block unit nodes; S102. Based on nighttime light data to characterize socioeconomic intensity and building volume to characterize physical carrying capacity, a two-factor proxy allocation model is constructed to calculate the carbon emission allocation weight of each block unit. S103. Calculate the true carbon emission label for each block unit based on the allocated weights and the total urban energy statistics.
3. The method for predicting urban carbon emissions using process feedback guidance and multi-scale graph neural networks according to claim 2, characterized in that, S2 include: S201. Extract the statistical features, texture features, and spectral index features of the street remote sensing image to form the image physical feature matrix; S202. Establish a multi-scale buffer with the street centroid as the origin, extract the interest point features within the buffer, and introduce a multi-head self-attention mechanism for feature encoding to generate functional semantic embeddings. S203. Surface temperature is retrieved using thermal infrared remote sensing data and used as an environmental feedback factor to characterize microclimate processes. S204. Utilize nighttime light data and demographic data to invert dynamic population density as an environmental feedback factor characterizing social activity processes.
4. The method for predicting urban carbon emissions using process feedback guidance and multi-scale graph neural networks according to claim 3, characterized in that, S3 include: S301. For any two street nodes, calculate their normalized absolute difference on the environmental feedback factor, construct an edge feature vector describing the environmental interaction relationship between nodes; and construct an environmental feedback graph tensor based on the edge feature vectors of all node pairs. S302. A multi-feature graph signal aggregation strategy is adopted, and a graph filter that depends on the environmental feedback graph tensor is used to dynamically generate the adjacency weights between nodes and update the node embedding. S303, Perform intermediate feature generation and aggregation.
5. The method for predicting urban carbon emissions using process feedback guidance and multi-scale graph neural networks according to claim 4, characterized in that, S4 include: S401. At the microscale, construct orientation-sensing nodes and distance-sensing nodes centered on ground features, and use graph convolutional networks to aggregate and obtain the street carbon spatial layout embedding. S402. At the macro scale, the feature representations of all blocks are aggregated using an area-weighted algorithm to generate a global city summary vector. S403. Construct positive and negative sample pairs by combining the street embedding representation with the city global summary vector, and maximize the mutual information between the two by minimizing the contrastive learning loss function.
6. The method for predicting urban carbon emissions using process feedback guidance and multi-scale graph neural networks according to claim 5, characterized in that, S5 include: S501. Construct positive sample pairs by combining the surface temperature map structure and dynamic population density map structure within the same block, and construct negative sample pairs by combining the surface temperature map structure and dynamic population density map structure between different blocks. S502. Construct a similarity measurement mechanism to evaluate the probability that a land surface temperature map and a dynamic population density map belong to the same spatial unit; S503. Construct an auxiliary contrast loss function based on the similarity measurement mechanism to bring positive sample pairs closer and negative sample pairs further apart in the feature space, thereby achieving collaborative constraints on heterogeneous modal data.
7. The method for predicting urban carbon emissions using process feedback guidance and multi-scale graph neural networks according to claim 4, characterized in that, S6 include: S601. Based on the fusion features obtained in the previous steps, obtain the carbon emission prediction value of the corresponding block; S602. Construct a multi-task joint loss function, which includes the regression loss of the main task and the contrastive learning loss of the auxiliary task. Learnable parameters are used to dynamically balance the gradient contributions of each task, and multi-task joint training is performed to optimize the model parameters.
8. A street-level carbon emission prediction system guided by process feedback and multi-scale graphical neural networks, characterized in that, For performing the method according to any one of claims 1 to 7, comprising: The topology segmentation module is used to perform topological segmentation of urban space using urban road network data to construct basic street block unit nodes, and to allocate the total urban energy statistics to each street block node based on proxy indicators to generate carbon emission training labels. The extraction module is used to extract the visual and functional semantic features of the block units and to invert the environmental feedback factors used to characterize microclimate processes and social activity processes. The generation module is used to dynamically generate edge weights between street nodes using the environmental feedback factor through a learnable graph filter, thereby constructing a dynamic street graph structure. A module is established to build a mutual information maximization model that includes multi-scale relationships. This model maximizes the mutual information between the street embedding representation and the city global embedding representation through a multi-scale contrastive learning task. The module is used to build a collaborative analysis model for heterogeneous multimodal data. It constructs positive and negative sample pairs from environmental feedback factors of different modalities. By maximizing the similarity of sample pairs within the same spatial unit and minimizing the similarity of sample pairs between different spatial units, it achieves multimodal information collaboration. The acquisition module is used to obtain carbon emission prediction results by taking the fusion features obtained from the construction module as input, combining the carbon emission numerical prediction error and the contrastive learning loss for multi-task joint training.
9. A computer storage medium, characterized in that, The computer storage medium stores a computer program; when the computer program is run on the computer, it causes the computer to perform the method described in any one of claims 1 to 7.
10. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the method as described in any one of claims 1 to 7.