Building embodied carbon assessment method based on multi-modal remote sensing and graph neural network

Abstract

This invention discloses a method for assessing building-embedded carbon emissions based on multimodal remote sensing and graph neural networks. The method includes: constructing a building structure type classification system for building-wide lifecycle carbon emissions; identifying building structure types and calculating lifecycle carbon emissions based on a building project inventory to generate a sample dataset; extracting and fusing building features based on multimodal remote sensing data to generate a unified building feature vector; jointly training a multi-task graph neural network using the sample dataset and the building feature vector to obtain a model capable of simultaneously predicting building structure type and lifecycle carbon emissions; and mapping and spatial analysis of building-wide lifecycle carbon emissions based on the model's prediction results. This invention achieves high-precision joint prediction of building structure identification and lifecycle carbon emissions, overcoming the limitations of traditional methods in data completeness and spatial accuracy, and enabling intelligent monitoring of urban carbon emissions.

Description

Technical Field

[0001] This invention belongs to the field of building carbon emission assessment technology, specifically involving a smart assessment technology solution for building structure and life-cycle implicit carbon based on multimodal remote sensing and graph neural networks. Background Technology

[0002] With global warming becoming increasingly severe, the construction industry, as a major source of energy consumption and carbon emissions (accounting for approximately 30% of global carbon emissions), has attracted widespread attention for its carbon reduction potential. Carbon emissions from the construction industry mainly consist of operational carbon and embodied carbon, with embodied carbon accounting for about 11% of total emissions, a proportion projected to rise to 50% by 2050. Embedded carbon emissions from buildings primarily originate from the lifecycle stages of building materials production, transportation, construction, replacement, and demolition / recycling. Compared to operational carbon, which is the direct energy consumption during the building's use phase, embodied carbon is more complex to calculate, exhibits greater spatial heterogeneity, and is difficult to obtain through conventional energy statistics or direct monitoring methods.

[0003] Existing methods for calculating implicit carbon throughout the building lifecycle mainly include top-down statistical methods and bottom-up spatial proxy methods.

[0004] Top-down statistical methods typically rely on macroeconomic statistics or industry carbon emission inventories, calculating carbon emissions from the construction industry through input-output models or allocation coefficients. This method has advantages such as a comprehensive calculation framework and broad coverage, reflecting overall emission levels at the national or regional scale. However, its limitations are also quite apparent: (1) The statistical or inventory data is not spatially complete, some indicators are missing, and it is difficult to achieve horizontal comparison between cities; (2) Statistical data are mostly released on an administrative district basis, which makes it difficult to depict the differences in carbon emissions of different building types within the administrative district; (3) The statistical data update cycle is long, mostly in the form of years, which cannot adapt to the rapidly changing needs of urban built environment monitoring.

[0005] Bottom-up spatial proxy methods use buildings or grid cells as analysis objects and leverage spatial proxy indicators such as nighttime light remote sensing, land use, or building density to estimate and visualize carbon emissions. This method offers high spatial resolution and visualization capabilities, but it still has the following limitations: (1) Spatial proxy data cannot directly reflect key influencing factors such as building structure type, material composition and service life, resulting in insufficient accuracy of implicit carbon estimation; (2) Nighttime light data mainly reflects the energy consumption intensity during the building operation phase, while the carbon emissions generated by building materials during the production and transportation phases are mostly located in suburban areas or off-site factories. The proxy variables and the actual emission points are spatially misaligned, making it impossible to achieve an accurate assessment of the entire building life cycle.

[0006] In summary, existing methods for calculating implicit carbon emissions throughout the building lifecycle have significant shortcomings in terms of data completeness, spatial accuracy, and building-level representation, making it difficult to support accurate estimation and mapping of implicit carbon emissions at the building level. Therefore, there is an urgent need to develop a refined assessment method that can combine building engineering data with multi-source remote sensing information to achieve spatialized and intelligent assessment of building structure identification and implicit carbon emissions throughout the building lifecycle. Summary of the Invention

[0007] To address the shortcomings of existing building lifecycle implicit carbon assessment methods, such as insufficient statistical data coverage, low spatial resolution, and misalignment between proxy data and actual building characteristics, this invention proposes a building implicit carbon assessment method based on multimodal remote sensing and graph neural networks. This method integrates building inventory data with multi-source remote sensing information to automatically identify building structure types and accurately estimate implicit carbon emissions. It also generates building lifecycle implicit carbon mapping results at a regional scale, thereby improving the spatial accuracy and automation level of building carbon accounting.

[0008] This invention addresses the aforementioned technical problems by providing a method for assessing building-borne carbon based on multimodal remote sensing and graph neural networks, comprising the following steps: Construct a classification system for building structure types that address the implicit carbon footprint throughout the entire building lifecycle; Based on the building engineering list, identify building structure types and calculate implicit carbon throughout the entire life cycle to generate a sample dataset; Building feature extraction and fusion are performed based on multimodal remote sensing data to generate a unified building feature vector; Based on a multi-task graph neural network, the sample dataset and the building feature vector are jointly trained to obtain a model that can simultaneously predict building structure type and implicit carbon emissions throughout the entire life cycle. Based on the prediction results of the model, mapping and spatial analysis of the implicit carbon throughout the building's entire life cycle are performed.

[0009] Furthermore, the building structure type classification system includes a preset set of building structure types, a set of building life cycle stages, and a set of building materials, and introduces an energy technology index to correct the carbon emission coefficient of materials for different regions.

[0010] Furthermore, the aforementioned identification of building structure types and calculation of implicit carbon throughout the entire life cycle based on the building project list specifically includes: Extract structural-related elements from the building project list to identify the building's structural type; The inventory data is categorized according to the preset life cycle stages, the implicit carbon emissions of each building material at each stage are calculated, and the total implicit carbon emissions of a single building throughout its entire life cycle are summarized to form a sample dataset.

[0011] Furthermore, the extraction and fusion of building features based on multimodal remote sensing data specifically includes: Acquire multi-source remote sensing data, including UAV oblique imagery, street view cloud data, and high-resolution satellite imagery; Registration and spatial alignment of multi-source data; Extracting architectural visual features from UAV imagery using a pre-trained visual model; The geometric and material features of facades in street view point clouds are extracted using a point cloud analysis model. Extracting spectral characteristics of building materials using high-resolution satellite imagery; The three features mentioned above are dynamically fused through an attention mechanism to generate a multimodal fusion feature vector.

[0012] Moreover, the multi-task graph neural network uses individual buildings as nodes and constructs a building relationship graph based on edge weights determined by the spatial distance between buildings, functional similarity, and differences in construction age. The node features are the multimodal fusion feature vectors. The network adopts a multi-layer graph sampling aggregation architecture, with each layer performing weighted sampling according to edge weights, and the number of samples gradually decreasing as the number of layers increases. The network shares parameters in the feature extraction layer, and the output layer is divided into two parallel branches, which are used for building structure type classification and full life cycle implicit carbon emission regression, respectively.

[0013] Moreover, the training of the multi-task graph neural network adopts a joint loss function, which includes structural classification loss, implicit carbon emission regression loss, structural smoothing loss for constraining the consistency of prediction of adjacent building structures, and physical law loss for constraining the positive correlation between carbon emissions and volume of the same type of building.

[0014] Furthermore, the cartography and spatial analysis specifically include: The model prediction results are mapped to geospatial data to establish a spatial database of building-hidden carbon based on individual buildings. Generate a regional-scale spatial distribution map of the implicit carbon footprint throughout the entire life cycle of buildings; Based on functional zoning or administrative boundaries, aggregate statistics and spatial autocorrelation analysis are conducted to identify the spatial clustering characteristics and distribution patterns of carbon emissions.

[0015] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the building implicit carbon assessment method based on multimodal remote sensing and graph neural networks as described above.

[0016] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the building implicit carbon assessment method based on multimodal remote sensing and graph neural networks as described above.

[0017] On the other hand, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the building implicit carbon assessment method based on multimodal remote sensing and graph neural networks as described above.

[0018] The above-mentioned technical solution addresses the problems of insufficient statistical data coverage, low spatial accuracy, and misalignment between proxy data and real building characteristics in existing building carbon assessments. It achieves high-precision joint prediction of building structure identification and life-cycle carbon emissions, breaking through the limitations of traditional methods in terms of data completeness and spatial accuracy. It can be widely applied to urban carbon emission monitoring and low-carbon building planning and implementation.

[0019] Compared with the prior art, the present invention has the following beneficial effects: (1) A unified building life cycle implicit carbon accounting system has been established.

[0020] This invention constructs a unified framework for building structure types, life cycle stages, and material carbon emission coefficients that address the implicit carbon emissions throughout the entire building life cycle. It clarifies the correspondence between building structure, material properties, and carbon emissions, and achieves standardization and systematization of the building implicit carbon calculation process, providing a consistent foundation for subsequent multi-source data fusion and model training.

[0021] (2) Integrate multimodal remote sensing information to improve the accuracy of building structure identification.

[0022] By comprehensively utilizing oblique UAV imagery, street view cloud data, and high-resolution satellite imagery, the system extracts building roof morphology, facade geometry, and spectral features, achieving multi-view, multi-dimensional representation of building characteristics. This multimodal fusion significantly improves the accuracy and stability of building structure identification in complex urban environments.

[0023] (3) Introduce graph neural networks to realize multi-task joint modeling.

[0024] This invention employs a graph neural network model based on a message-passing mechanism, capable of simultaneously performing building structure classification and life-cycle implicit carbon emission regression tasks. The model shares semantic information at the feature layer and performs task branch optimization at the output layer, fully considering spatial proximity, functional correlation, and age differences between buildings to achieve collaborative optimization of structure identification and implicit carbon estimation.

[0025] (4) Realize the implicit carbon mapping and spatial analysis of buildings at the regional scale.

[0026] This invention maps model inference results to a geospatial database, using individual buildings as the smallest spatial unit to achieve building-level implicit carbon mapping. Combined with global and local spatial autocorrelation analysis, it can identify high-emission clusters and spatial anomaly areas, quantitatively revealing the spatial distribution patterns of implicit carbon in regional buildings, and providing a scientific basis for urban building carbon emission monitoring, low-carbon design, and emission reduction planning. Attached Figure Description

[0027] Figure 1 This is a general framework diagram of the method of the present invention.

[0028] Figure 2 This is a schematic diagram of a building relationship diagram constructed using buildings as nodes and spatial and semantic associations as edges, according to an embodiment of the present invention.

[0029] Figure 3 This is a schematic diagram of the architecture of the building structure classification and implicit carbon emission joint prediction model based on a multi-task graph neural network according to an embodiment of the present invention.

[0030] Figure 4 This is an example of a spatial distribution map showing the predicted carbon emissions throughout the entire life cycle of buildings in a region, according to an embodiment of the present invention. Detailed Implementation

[0031] To make the technical solution of the present invention clearer and easier to understand, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments are only used to illustrate the present invention and are not intended to limit the scope of protection of the present invention.

[0032] Example 1 A method for assessing building-embedded carbon based on multimodal remote sensing and graph neural networks includes the following steps: Step 1: Construct a building structure classification system for implicit carbon emissions throughout the entire building lifecycle; Specifically, the set of building structure types, life cycle stages, and main building materials is determined, and the benchmark strength and carbon emission coefficient of the materials are established to provide a standardized basis for subsequent calculations.

[0033] Step 2: Identify building structures and calculate implicit carbon throughout the entire life cycle based on the building project inventory to generate a sample dataset; Specifically, information on building structure type and material composition is extracted from the building project list data, and the implicit carbon emissions under different structure types and life cycle stages are calculated to form a true value dataset of building samples.

[0034] Step 3: Extract and fuse building features based on multimodal remote sensing data to generate a unified building feature vector; Specifically, the system extracts architectural visual features, facade geometric and material features, and spectral features from UAV imagery, street view cloud images, and high-resolution satellite imagery, respectively. Through cross-modal attention, it achieves weighted fusion of multimodal features and ultimately generates a unified set of architectural description feature vectors.

[0035] Step 4: Based on the multi-task graph neural network, the sample dataset and the building feature vector are jointly trained to obtain a model that can simultaneously predict the building structure type and the implicit carbon emissions throughout the entire life cycle. Specifically, a graph neural network model is constructed using individual buildings as nodes and the spatial adjacency relationships between buildings as edges. Multi-task learning is performed using node feature vectors to achieve joint prediction of building structure classification and implicit carbon emissions throughout the entire life cycle. Physical consistency constraints and structural smoothing regularization terms are introduced during model training to improve prediction stability and physical rationality.

[0036] Step 5: Based on the prediction results of the model, perform mapping and spatial analysis of the implicit carbon throughout the building's life cycle.

[0037] Specifically, the prediction results are mapped to geospatial data to generate a city-scale map of the implicit carbon emissions throughout the building lifecycle. Using functional zoning or administrative boundary data as the zoning statistical boundaries, the total amount and intensity of implicit carbon emissions throughout the building lifecycle at the regional scale are statistically analyzed. Spatial autocorrelation indicators such as Global Moran's I and Local Moran's I are used to analyze the spatial clustering characteristics and distribution patterns of implicit carbon emissions throughout the building lifecycle.

[0038] Example 2 The overall framework diagram of the building implicit carbon assessment method based on multimodal remote sensing and graph neural networks proposed in this invention is as follows: Figure 1 As shown. The specific implementation method mainly includes the following steps: S1: Construction of a building structure classification system for implicit carbon throughout the entire building life cycle; In one possible embodiment, step S1 is preferably implemented using the following sub-steps: S1.1: Define the set of building structure type classifications as follows .

[0039] In one possible embodiment, a set of classifications for building structure types is determined. Among them, the preferred recommendation is... It is a brick-concrete structure. It is a framework structure. It is a shear wall structure, It is a steel structure. It has a prefabricated structure and can be expanded further. It represents the total number of building structure types; This invention proposes defining a set of building structure type classifications as follows: Building structure mainly refers to the main composition of the load-bearing framework of a building. Different building structures differ significantly in load-bearing type, construction technology, and main materials. For example, timber structures use wood as the main load-bearing material and are commonly found in ancient buildings and lightweight houses. Brick-concrete structures use walls for load-bearing and reinforced concrete beams and slabs, and are common in multi-story residential structures. Steel structures are often used in industrial plants, bridges, and large-span buildings. Based on the main force system and main material source of the building structure, classifications include: timber structures, brick and stone structures, brick-concrete structures, reinforced concrete structures, frame structures, shear wall structures, steel structures, steel-concrete composite structures, prefabricated structures, and hybrid structures. Specific classifications can be adaptively expanded according to regional characteristics. For example, for northern cities, light steel structures can be added; for southern regions, bamboo structures or block structures can be added. This invention does not limit these types of classifications.

[0040] S1.2: Determine the set of building life cycle stages Preferably, it includes five stages: building material production, construction, use, end-of-life, and supplementary information beyond the system boundaries. It can also be expanded to include more stages. It is the total number of stages; Implementation examples determine the set of building lifecycle stages ,in, For the material production stage, For the construction phase, For operation and maintenance phase (use) This is the dismantling phase (end of life). For the material recycling phase (supplementary information outside the system boundaries); This approach is based on the building lifecycle phase division framework specified in ISO 21930, defining the set of building lifecycle phases as follows: It includes five stages: building material production, building construction, building operation and maintenance, building demolition, and material recycling. This stage division directly corresponds to the system boundary setting in the calculation of implicit carbon emissions throughout the entire life cycle.

[0041] S1.3: Determine the main building materials Preferably, It is cement, It is sand; other materials will be determined based on specific circumstances. It refers to the quantity of the main building materials.

[0042] In this embodiment, to establish the correspondence between the material characteristics of different building structures and carbon emission calculations, a set of main building materials is established. This includes, but is not limited to, cement, sand, stone, brick, steel reinforcement, glass, wood, aluminum, copper, and lime. For each material, a baseline carbon emission factor is defined. This represents the carbon emissions per unit of material production process. Data is primarily sourced from carbon emission factors provided by upstream material suppliers; when these are unavailable, the "Building Carbon Emission Calculation Standard" or the IPCC default emission factor is used. To supplement.

[0043] To further adapt to regional characteristics, a material correction factor is introduced. At the provincial level, the default carbon emission factor The revisions are intended to reflect regional differences in energy structure. Specifically, they are based on the assumption that a region with a cleaner energy structure, more advanced technology, and higher industrial energy efficiency will have lower emissions from materials production. The proportion of non-fossil energy power generation in the region is obtained from national and provincial energy statistical yearbooks and statistical bulletins on national economic and social development. Energy consumption per unit of industrial added value Regional per capita GDP These three indicators represent the region's energy cleanliness, industrial energy efficiency, and technological and economic development level, respectively. The three indicators are normalized and their values ​​range from [0,1]. The regional energy technology index... It can be represented as:

[0044] Based on this, the material correction factor It can be represented as:

[0045] in, It represents the national average energy technology index.

[0046] Ultimately, in the absence of carbon emission factors provided by upstream material suppliers, the benchmark carbon emission coefficient for a certain material in the region... It can be represented as:

[0047] in, This is the default carbon emission factor for the IPCC.

[0048] S2: Building structure identification and life-cycle implicit carbon calculation based on the building project list; Specifically, S2 only performs statistical calculations on buildings within the region that have complete building project list data. Complete building project list data must provide at least three key parameters: building material usage, energy consumption, and material production location.

[0049] In one possible embodiment, step S2 is implemented using the following sub-steps: S2.1: Extract structural-related elements from the building bill of quantities, including material type, number of components, construction procedures, and equipment energy consumption information; Extract elements related to building structure and carbon emissions from the building bill of quantities, including material type, component quantity, construction sequence, and equipment energy consumption information. Material type information is used to identify the main structural materials of the building, such as concrete, steel bars, bricks, glass, and wood; component quantity and specifications reflect the scale of the building structure and the proportion of materials used; construction sequence information (such as pouring, welding, and assembly) can be used to identify the characteristics of the building structure type; energy consumption information includes electricity and fuel consumption during the construction phase, used for subsequent carbon emission calculations.

[0050] S2.2: Based on the extracted building structural elements and building structural classification set The correspondence is used to identify the structural type of the building; Based on the extracted building structural elements and the pre-established set of building structural classifications The system performs matching to identify the building's structural type. This information is primarily derived from relevant records of the building structure in the bill of quantities, and secondarily from a comprehensive judgment method based on material proportions and construction procedures: if concrete and steel reinforcement constitute the majority of the total material, it is identified as a "frame structure"; if bricks and mortar account for more than 60% of the walls, it is identified as a "brick-concrete structure"; if steel components are significant in number, and the main components are "H-beams," "steel beams," and "steel columns," it is identified as a "steel structure"; if concrete components include a large proportion of shear walls, it is identified as a "shear wall structure"; if the bill of quantities contains a large number of precast slabs, precast walls, or connectors, it is identified as a "prefabricated structure."

[0051] S2.3: Based on the aforementioned lifecycle stage set The inventory data is categorized by stage to calculate the implicit carbon emissions of specific building materials in individual buildings at different stages of their life cycle.

[0052] In the embodiment, based on the lifecycle stage set This study calculates the hidden carbon emissions from buildings across five stages: building material production, construction, operation and maintenance, demolition, and material recycling. For the material production stage, hidden carbon emissions primarily originate from the extraction and processing of building materials. The calculation formula is as follows:

[0053] in, It is a building material Hidden carbon emissions during the materials production stage, This refers to the total amount of building materials used during the construction phase, which can be obtained from the building bill of quantities. It is the carbon emission intensity coefficient of the building material.

[0054] During the construction phase, the hidden carbon emissions of a building mainly come from the transportation of materials from mining and processing plants to the construction site. The calculation formula is as follows:

[0055] in, It is a building material Hidden carbon emissions during the building construction phase Building materials during the construction phase Average transport distance This refers to the carbon emission intensity from transporting building materials. The value can be determined by referring to the building carbon emission calculation standard issued by the Ministry of Housing and Urban-Rural Development of the People's Republic of China.

[0056] During the building operation and maintenance phase, the building's implicit carbon emissions mainly come from the maintenance and replacement of building components, as calculated using the following formula:

[0057] in, It is a building material Hidden carbon emissions during the building operation phase This refers to the annual maintenance and renewal rate of a building. As a building ages, both maintenance and renewal rates gradually increase, with an average maintenance rate ranging from 0.3% to 2.8%. Maintenance rates vary slightly depending on the building structure: brick-concrete structures have relatively weaker aging resistance, and problems such as mortar grouting, wall cracking, and water seepage are more common, with an average maintenance rate of 1.5% to 2.5%; reinforced concrete structures have a maintenance rate between 1% and 2%; and steel structures have a maintenance rate between 0.5% and 2.5%. It is the carbon emission intensity coefficient of the building material.

[0058] During the building demolition phase, the main source of hidden carbon emissions from buildings is waste transportation, calculated using the following formula:

[0059] in, It is a building material Hidden carbon emissions during building demolition Building materials for the demolition phase The average transport distance is 50 km for recyclable materials such as steel, sand, and gravel, and 30 km for other non-recyclable materials. This refers to the carbon emission intensity of transportation, and the value can be determined by referring to the building carbon emission calculation standards published by the industry.

[0060] During the material recycling stage, the hidden carbon emissions from buildings mainly come from the further processing of recyclable materials, as calculated using the following formula:

[0061] in, It is a building material Hidden carbon emissions during the post-recycling and processing stage of materials, For building materials The recyclability rates are 50%, 30%, and 30% for steel, sand, and gravel, respectively. This represents the carbon intensity during the processing of recyclable materials, typically 105% of the baseline carbon intensity.

[0062] Finally, building materials Total carbon emissions hidden throughout the entire life cycle Calculate using the following formula:

[0063] S2.4: Construct a sample set of implicit carbon emissions throughout the entire life cycle of a single building.

[0064] For buildings with building construction inventory records, their building structure type and total usage of all building materials are obtained. Step S2.3 is then used to calculate the total occult carbon emissions of each type of building material throughout its entire lifecycle, and these values ​​are summed to obtain the occult carbon emissions of a single building throughout its entire lifecycle. :

[0065] Building structure types obtained from engineering bill of quantities data and the hidden carbon emissions throughout the entire life cycle of a single building The set of true samples is used as training data for the graph neural network model.

[0066] S3: Building feature extraction based on multimodal remote sensing data; This step further includes: Acquire oblique drone imagery, street view cloud data, and high-resolution satellite imagery; Register and spatially align multi-source data to establish a unified geographic coordinate reference framework; Extracting architectural visual features from UAV imagery using a pre-trained visual model; The geometric and material features of facades in street view point clouds are extracted using a point cloud analysis model. Extracting spectral features of buildings using high-resolution satellite imagery; The visual features, facade geometry and material features, and spectral features are weighted and fused using an attention mechanism to form the unified architectural feature vector.

[0067] In one possible embodiment, S3 is preferably implemented using the following sub-steps: S3.1: Acquire multi-source remote sensing data, including UAV oblique imagery, street view cloud data, and high-resolution satellite imagery; The implementation plan targets the study area, utilizing a small UAV equipped with a multi-lens oblique photography system to conduct aerial photography under flight conditions meeting the requirements of ≥80% forward overlap and ≥70% lateral overlap, acquiring oblique images from the UAV. A ground-based laser scanning system was used to collect high-density point clouds of buildings on both sides of the street, with a point density of 500–1000 points / square meter. The collected point cloud data underwent coordinate system optimization, noise filtering, and feature segmentation, preserving geometric features such as building facades, windows, walls, and structural components, serving as street view point cloud data. GF-2 data with a spatial resolution better than 1 meter, covering four bands—red, green, blue, and near-infrared—was selected as high-resolution satellite imagery data.

[0068] S3.2: Register and spatially align multi-source data to establish a unified geographic coordinate reference framework; In this embodiment, the multi-source data includes UAV oblique imagery, street view point cloud data, and high-resolution satellite imagery. Aerial triangulation and ground control point calculation are performed on the UAV oblique imagery to generate orthophotos. Point cloud data is spatially aligned with the DOM (Domain of Imagery) using a feature point matching method. Satellite imagery is radiometrically calibrated and atmospherically corrected, and then unified to the WGS84 geographic coordinate system using multi-control point registration. Finally, the three multimodal data are unified into a single geographic coordinate reference frame, ensuring that the geometric overlap error of the multi-source data at the building boundary scale does not exceed 0.2 meters. Data is sliced ​​using individual buildings as the smallest spatial unit, and the sliced ​​UAV imagery, street view point cloud data, and satellite multispectral data correspond to the same building entity.

[0069] S3.3: Extract architectural visual features from UAV imagery using a pre-trained visual model; The embodiment inputs sliced ​​UAV image blocks into a pre-trained visual model to obtain building visual feature vectors. In practice, classic visual deep learning models such as Swin transformer, RseNet, and CLIP can be used to extract the building's roof shape, spatial texture, and spectral features, ultimately outputting a multi-dimensional visual feature vector of the building. .

[0070] In this embodiment, the Swin transformer visual model architecture is preferably introduced. For specific implementation, please refer to the paper "Swin Transformer: Hierarchical Vision Transformer using Shifted Windows" published by Liu et al. in 2021. For ease of implementation, the following specific description is provided: The model first divides the image block into q non-overlapping image blocks through a slicing module. Then, through a linear embedding module, each image block is mapped to a d-dimensional feature vector to form an initial feature map. The initial feature map is processed in the following stages four times: each stage includes a window-based multi-head self-attention module, a shifted window-based multi-head self-attention module, and a channel downsampling module. These three modules are alternately connected to achieve comprehensive feature capture of architectural images from fine-grained texture to macroscopic geometric structure. The window-based multi-head self-attention module aims to accurately capture detailed information such as building wall material and surface texture. The shifted window-based multi-head self-attention module aims to establish cross-window feature dependencies and identify long-range dependencies of the building, such as macroscopic geometric shape, edge contour structure, and spatial layout information. The channel downsampling module halves the feature map size and doubles its dimension to progressively improve the semantic meaning and global relevance of the features. The features, after four iterations, are finally mapped through a fully connected layer to obtain a 64×1×1 architectural visual feature vector. This embodiment has been experimentally verified to optimally balance data information abundance and computational efficiency with a feature dimension of 64×1×1.

[0071] S3.4: Extract the facade geometry and material features from street point cloud using a point cloud analysis model; The example inputs the sliced ​​street view cloud data into a point cloud analysis model to obtain building facade feature vectors. Point cloud analysis models such as PointNet++ and KPConv can be used to extract facade normal vector distribution, window-to-wall ratio, component nodes, and surface texture features, ultimately outputting multi-dimensional geometric and material feature vectors of the building facade. .

[0072] In this embodiment, the PointNet++ point cloud processing architecture is preferably introduced. For specific implementation details, please refer to the paper "PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation" published by Qi et al. in 2017. For ease of implementation, the following explanation is provided: Geometric features such as facade normal vector distribution, window-to-wall ratio, surface roughness, and node density are automatically extracted through point neighborhood convolution and normal estimation. Simultaneously, the facade material features are estimated by combining point cloud reflection intensity. Finally, a fully connected layer is used to map and obtain a 64×1×1 feature vector reflecting the geometry and material of the building facade, denoted as the point cloud facade feature. This embodiment has been experimentally verified to optimally balance data information abundance and computational efficiency.

[0073] S3.5: Extracting spectral characteristics of building materials using high-resolution satellite imagery; Unlike conventional methods that extract the overall spectral characteristics of a building, this step focuses on extracting the spectral characteristics of building materials. This is because the occult carbon emissions of a building are directly related to the type of building materials (such as concrete, steel, and bricks), and different materials have unique spectral response characteristics in satellite imagery. By extracting the spectral characteristics of building materials, the proportion of materials used in a building can be more accurately inferred, thereby improving the accuracy of occult carbon assessment.

[0074] The embodiment calculates the surface reflectance of sliced ​​high-resolution satellite images in red, green, blue, near-infrared, and short-wave infrared bands, and extracts brightness temperature features in the presence of thermal infrared bands to characterize the optical and thermal differences of building surfaces. To comprehensively reflect the spectral and energy absorption characteristics of building materials, the Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Normalized Difference Soil Index (NDSI), and Normalized Difference Building Index (NDBI) are further calculated. Simultaneously, image texture features are extracted based on the Gray-Level Co-occurrence Matrix (GLCM) to comprehensively characterize the reflectance and energy absorption capacity of building materials. Principal component analysis yields a 64×1×1 satellite feature vector reflecting the spectral characteristics of building materials. This embodiment preferably uses a 64×1×1 feature dimension, which can ensure the accuracy of building structure recognition while taking into account the computational efficiency of the subsequent graph neural network. Experimental verification has shown that this optimal balance between data information abundance and computational efficiency can be achieved.

[0075] S3.6: Integrate different modal features to form a descriptive feature vector set for implicit carbon throughout the entire building life cycle.

[0076] In practical implementation, the visual features, facade geometry and material features, and spectral features can be weighted and fused using an attention mechanism to form the unified architectural feature vector, including: By combining embedding alignment and feature concatenation, and using cross-modal attention to achieve weighted fusion of multimodal features, a unified 64-dimensional building feature vector is generated.

[0077] in, , , Representing buildings The 64-dimensional visual features, 64-dimensional facade geometric and material features, and remote sensing features.

[0078] The example uses the visual features of the drone Point cloud facade features and spectral characteristics of building materials The input feature fusion module generates feature description vectors for individual buildings. This invention further proposes addressing the issue that traditional direct stitching of data from three different modalities—UAV visual features, point cloud facade features, and building material spectral features—exhibits differences in data distribution and semantic information, making effective interaction difficult. Therefore, steps S3.3, S3.4, and S3.5 project the heterogeneous data into a high-dimensional semantic subspace of the same dimension, followed by dynamic feature fusion through a cascaded attention and gating mechanism. First, self-attention operations are performed independently for each modality to achieve feature purification and denoising. This is based on the visual features of drones. For example, three different feature vectors are generated: ,in, , , These represent three learnable weight matrices, which are continuously updated during network training. According to the formula... Calculate the self-attention weights ,in represent transpose, for Dimensions It is a normalization operation that normalizes the weights. With original features The summation and layer normalization yield new features with self-attention. Similarly, the point cloud facade features are obtained. and satellite spectral characteristics The corresponding new features , .

[0079] Secondly, multimodal information weighted fusion is achieved through a cross-modal attention mechanism: Sum and average the results, then input them into a fully connected layer to generate a global query vector. ,Will Split by dimension The features of each group are analyzed, and single-head attention is calculated for each group. Taking a group as an example, the features of the three modalities are concatenated to obtain: No. Group key vector , No. Value vector of a group , in, and These represent two learnable weight matrices, and are respectively connected to... Interaction, to obtain the first Feature attention of groups Next will The attention results of the sub-features are concatenated to obtain the cross-attention features. ,in It is a learnable linear projection matrix.

[0080] Furthermore, a gating mechanism is introduced for each modality to dynamically control the contribution of each modality. Again, taking UAV visual features as an example, the modal contribution... ,in Activation function , It is a multilayer perceptron, and similarly, the point cloud facade features are obtained. and satellite spectral characteristics The respective contribution Weighted fusion modal features Finally, the cross-attention features Weighted modal features The features are spliced ​​together and fused through a fully connected layer to obtain multimodal fusion features. This feature integrates multi-dimensional information such as the building's roof shape, facade geometry, and material spectrum, and will serve as the initial input feature for the corresponding building nodes in the subsequent graph neural network model.

[0081] S4: A joint training model for building structure and implicit carbon based on a multi-task graph neural network; This step further includes: Using buildings as nodes and the spatial, functional, and chronological relationships between buildings as edges, construct an architectural relationship diagram; The building feature vector is used as the feature of the corresponding node and input into the multi-task graph neural network; The multi-task graph neural network is trained using the designed joint function; Output the prediction results of building structure type and the estimated value of implicit carbon emissions throughout the entire life cycle.

[0082] In one possible embodiment, S4 is preferably implemented using the following sub-steps: S4.1: Construct an architectural relationship diagram with buildings as nodes and the spatial, functional, and chronological relationships between buildings as edges; When constructing the building relationship graph, the weight of the edge is determined based on the spatial distance between the corresponding two buildings, the functional similarity, and the difference in construction years.

[0083] In this embodiment, each building within the study area is considered as a node in the graph, and the set of nodes is denoted as . Each node It contains one input: a feature vector. And two outputs: building structure type and hidden carbon emissions throughout the entire life cycle The set of edges between nodes Representation Nodes and The spatial and semantic relationships between nodes. This invention proposes that edge weights be determined based on the nodes. and spatial distance Functional similarity and differences in construction period The determination is based on a comprehensive approach. First, spatial distance and temporal differences are normalized using the maximum value: Let the maximum shortest road network distance between any two buildings within the study area be... Then the normalized spatial distance Let the maximum difference in the completion year between the buildings be _____. Then normalize the differences in construction dates. Then, based on functional similarity, the types and quantities of POIs within a building are counted to form an n-dimensional vector. The cosine similarity between the two buildings is then calculated as the functional similarity. The representation is then used. The edge weights are calculated using a weighted product model, as shown in the following formula:

[0084] in, The attenuation coefficient is the spatial distance. The larger the value, the faster the weight decays as the distance increases. The importance weight coefficients for each factor can be determined based on the Analytic Hierarchy Process (AHP) or the actual application scenario.

[0085] The edge weights comprehensively reflect the spatial proximity, functional similarity, and consistency of construction dates among buildings. Specifically, the spatial distance attenuation term reflects that buildings closer together are more likely to have similar structural types and material compositions; functional similarity is based on the engineering principle that buildings of the same functional type (such as residential, commercial, and industrial) often adopt similar structural designs; and the construction date difference term reflects the impact of the evolution of building codes and technical standards in different periods on building structures. By integrating these three factors through a weighted product model, a graph structure that truly reflects the internal relationships within a building cluster is constructed, providing neighborhood relationships consistent with engineering realities for subsequent graph neural network information aggregation.

[0086] like Figure 2 As shown, taking four individual buildings as an example, the final node relationship diagram is obtained as follows: , , These represent the edge weights between building 1 and building 2, building 3, and building 4, respectively. , These represent the edge weights between building 2 and building 3, and building 4, respectively. This represents the edge weight between building 3 and building 4.

[0087] This building relationship graph constructs the topological structure between building nodes, providing a foundation for subsequent neighborhood sampling and information aggregation in the GraphSAGE layer. During the training of the graph neural network, nodes will aggregate feature information of neighboring buildings through an edge-weighted sampling strategy, thereby achieving joint prediction of building structure type and implicit carbon emissions.

[0088] S4.2: Input the building feature vector as the feature of the corresponding node into the multi-task graph neural network; In this embodiment, a two-layer GraphSAGE is used as the core, and a dual-branch graph neural network model based on edge weights is designed by combining the characteristics of building node features and edge features. In the existing graph neural network aggregation process, random sampling of nodes is usually used to aggregate information, which can greatly reduce computational costs, but it is easy to introduce noisy nodes and dilute high-value information. To address this problem, this invention proposes to adopt a sampling mechanism based on node similarity. The principle is to use the similarity between nodes to guide the sampling and information weighting aggregation process, so that the model pays more attention to those nodes that are closer to the central node in semantic space. The purpose of this improvement is to: (1) efficiently filter topological noise in the graph structure; (2) ensure that the aggregated feature vector can retain the essential attributes of the nodes to the greatest extent; (3) improve the feature expression ability of the model without significantly increasing computational complexity. In addition, based on the prior knowledge that building structure information and building implicit carbon emissions are strongly correlated, a dual-branch network is introduced. Branch 1 is for building structure prediction, and branch 2 is for full life cycle implicit carbon emission prediction. The two share the underlying feature extraction. See Figure 3 The network mainly consists of an input layer, a first-order GraphSAGE layer, a second-order GraphSAGE layer, and an output layer. With two GraphSAGE layers at its core, the network employs a weighted sampling strategy based on edge weights and a dual-branch output architecture to achieve joint prediction of building structure type and implicit carbon emissions throughout the entire life cycle. Specific details are as follows: 1) In the input layer, each building node initial features It is the multimodal fusion feature vector obtained in step S3.6 And based on edge weights Constructing an adjacency matrix It is used for neighborhood sampling and aggregation.

[0089] 2) In the first layer GraphSAGE, for each node First, the sampled layer is weighted. Unlike traditional GraphSAGE, this implementation improves uniform sampling to edge-weighted sampling, that is, for nodes... First-order neighborhood set Calculate neighbors sampling probability This emphasizes that neighbors with larger edge weights have a higher probability of being sampled, guiding information interaction between highly correlated nodes. Based on network prediction accuracy and speed, an optimal sampling number is set using a grid search method. The set of sampling points is obtained. In the linear mapping layer, for nodes The initial features and all its sampled neighbor nodes The initial features are linearly mapped to obtain node features. The corresponding formula is ,in, and Representing two learnable parameters; in the feature weighted aggregation layer, the mean aggregation is improved to neighborhood feature weighted aggregation according to edge weights, resulting in neighborhood aggregated features. The corresponding formula is In the feature activation layer, node features are... Features of neighborhood aggregation The nodes are concatenated, then pass through a linear layer and an activation layer to output nodes. Node features after the first layer of GraphSAGE .

[0090] 3) To capture semantic relationships at the building cluster level, in the second layer of GraphSAGE, for nodes output from the first layer... feature In the weighted sampling layer, based on the network's prediction accuracy and speed, a grid search method is used to search its first-order neighbor set. Weighted sampling by edge weight 15 nodes, i.e., nodes First-order neighborhood set Calculate neighbors sampling probability It emphasizes that neighbors with larger edge weights have a higher probability of being sampled, guiding information interaction between highly correlated nodes, and then in each sampled... First-order neighbor set Weighted sampling based on edge weights 10 nodes, i.e., a pair of nodes First-order neighborhood set Calculate neighbors sampling probability It emphasizes that neighbors with larger edge weights have a higher probability of being sampled, guiding information interaction between highly correlated nodes and forming nodes. Second-order weighted sampling neighborhood And repeat the first layer structure through the linear mapping layer ( ,in, and Represents two learnable parameters. For nodes The output after the linear mapping layer), the feature weighted fusion layer ( ,in, (representing aggregated features of first- and second-order neighborhoods) and feature activation layers (concatenation) and After passing through a linear layer and an activation layer, the output yields the second layer of node features. .

[0091] 4) Furthermore, based on the features of the second layer output It is divided into two parallel branches: Branch 1 performs building structure category prediction, which includes a first fully connected layer, a first nonlinear activation layer (ReLU), a second fully connected layer, and a second nonlinear activation layer connected in sequence. This branch aims to process the input feature vector The data is mapped to category logical values ​​and transformed into predicted probability distributions for each category through a Softmax layer to classify building structures (such as brick-concrete structures, frame structures, shear wall structures, steel structures, prefabricated structures, etc.). Branch 2 performs full life-cycle implicit carbon emission prediction. Its network structure is similar to that of branch 1, containing alternating fully connected layers and activation layers. This branch aims to establish a numerical regression model between feature vectors and implicit carbon emissions, directly outputting the predicted values ​​of full life-cycle implicit carbon emissions.

[0092] This dual-branch network architecture shares the underlying GraphSAGE feature extraction layer, enabling the model to fully leverage the strong correlation between building structure and implicit carbon emissions during the learning process. Different structural types of buildings have significantly different material compositions and carbon emission characteristics. Sharing feature extraction promotes mutual enhancement between the two tasks, improving prediction accuracy. It should be noted that the sampling numbers of 20, 15, and 10 mentioned above are optimal suggested values ​​obtained through grid search based on network prediction accuracy and speed. In practical applications, these values ​​can be adjusted according to the specific task and data scale, but should follow the principle of gradual reduction to balance information utilization and computational efficiency. The training process of the multi-task graph neural network will employ a weighted sampling strategy based on edge weights to aggregate neighborhood node information.

[0093] S4.3: Train the multi-task graph neural network using the designed joint function; The loss function of the joint training includes structure classification loss, implicit carbon emission regression loss, structure smoothing loss to constrain the consistency of predictions for adjacent building structures, and physical law loss to constrain the positive correlation between carbon emissions and volume of the same type of building.

[0094] In this embodiment, the network is trained using a joint loss function, which consists of the following losses: (1) The structural classification loss is defined as the difference between the network-predicted building structure and the actual building structure, and the calculation formula is as follows:

[0095] in, It is structural classification loss. For building sample size, For the number of building structure categories, Indicates the first The building node sample belongs to the first The true labels for each structural category (using one-hot encoding format; 1 if the sample belongs to that category, 0 otherwise). For network for building nodes Belongs to the building structure type The predicted probability, the structural classification loss is used to constrain the difference between the network's predicted probability of building structure type and the true label, and to ensure the accuracy of structural classification.

[0096] (2) The regression loss of implicit carbon emissions is defined as the root mean square error between the actual implicit carbon emissions result and the predicted result. The calculation formula is as follows:

[0097] in, It is the implicit carbon emission regression loss. For building sample size, It is a building The true, implicit carbon emission structure, It is the network for buildings The implicit carbon emission prediction results.

[0098] (3) According to the first law of geography, adjacent buildings in a city usually have consistent structural types. By penalizing the differences in structural type predictions between adjacent nodes, the network's ability to capture the structural patterns of building clusters is improved. The formula for calculating the structural consistency constraint loss is as follows:

[0099] in, It is the structural consistency constraint loss. For building sample size, It is a node The first-order sample set, Represents a set The number of elements, It is a node and Similarity weights between them These represent network connections to building nodes. and building nodes Distribution vector of predicted probability for building structure type. This represents the KL divergence, which is used to measure neighboring nodes. Predicted distribution and central node The difference between the predicted distributions.

[0100] (4) According to engineering principles, for buildings of the same structural type, the larger the volume, the greater the implicit carbon emissions over the entire life cycle. A volume-carbon emission correlation constraint is set, and the calculation formula is as follows:

[0101] in, It is a volume-carbon emission-related constraint loss. It refers to the number of building structure types. Indicates belonging to the first A collection of architectural samples of various architectural structure types. Indicates the number of elements in the set. and These represent the model's connection to building nodes. and building nodes The implied carbon emission forecast. and Representing building nodes and building nodes The building volume. Represents a sign function, when When the value is 1, The value is -1 when When the value is 0, it indicates the building node. and building nodes The relative volume size. This represents hinge loss, which only occurs when the relative magnitudes of the predicted values ​​contradict the relative magnitudes of the volumes.

[0102] Based on the above losses, the optimal weights for each loss are determined through grid search and then summed. In this embodiment, the weight coefficients of each loss are optimized on the validation set using the grid search method, and the final weight combination is determined as follows: structural classification loss weight 1.0, implicit carbon emission regression loss weight 1.2, structural consistency constraint loss weight 0.5, and volume-carbon emission correlation constraint loss weight 0.7. This combination can effectively introduce physical knowledge constraints while ensuring basic prediction accuracy, avoiding the suppression of other tasks' learning by one dominant task. Joint Loss The calculation formula is as follows:

[0103] The joint loss function above combines the loss due to data fitting accuracy and physical knowledge constraints. and Ensure the network's basic fitting accuracy to the real labels; and By introducing spatial consistency from the first law of geography and volume monotonicity from building physics as physical constraints, the solution space for multi-task learning is effectively reduced, preventing the network from generating predictions that violate engineering common sense. A grid search strategy is used to determine weight coefficients, balancing the magnitude differences in gradients of different loss functions and preventing a dominant task from inhibiting the learning of other auxiliary tasks, thereby achieving multi-objective collaborative convergence.

[0104] The training samples are derived from the ground-value sample set of building structures and their implicit carbon emissions throughout their entire lifecycle, generated from the engineering inventory data in step S2. The network is trained using the Adam optimizer, with an initial learning rate set to [value missing]. The batch size is 128. To prevent overfitting, a Dropout mechanism is introduced between network layers, and the validation set error is monitored using an Early Stopping strategy.

[0105] S4.4: Output the prediction results of building structure type and the estimated value of implicit carbon emissions throughout the entire life cycle.

[0106] In this embodiment, after the network is trained, it infers the building structure and life-cycle implicit carbon emissions of all buildings in the region, and outputs the structure type of each building and its corresponding life-cycle implicit carbon emissions.

[0107] S5: Hidden Carbon Mapping and Results Analysis of Building Life Cycle The mapping and spatial analysis of implicit carbon throughout the building life cycle includes: mapping the predicted building structure type and implicit carbon emissions to geospatial space, generating a building-level carbon emission distribution map, and performing spatial autocorrelation analysis to identify spatial clustering patterns of carbon emissions.

[0108] In one possible embodiment, S5 is preferably implemented using the following sub-steps: S5.1: Map the building structure type and implicit carbon emission results predicted by the graph neural network to geospatial data; The example uses a trained graph neural network model to perform inference calculations on all buildings within the study area, obtaining the structural type of each building and its corresponding life-cycle implicit carbon emissions. The network prediction results are spatially paired with the building vector boundaries using the building's unique spatial identifier and geographic coordinates. After pairing, each building entity contains three types of attribute information: building structural type, total life-cycle implicit carbon emissions, and carbon emission intensity per unit area, achieving a spatial correspondence from numerical output to geographic entities. This spatial mapping result can serve as the data foundation for subsequent database construction.

[0109] S5.2: Establish a building-hidden carbon space database using individual buildings as the basic spatial unit; This example uses individual buildings as the smallest spatial unit and constructs a database of implicit carbon emissions for each building in PostgreSQL. The database fields mainly include building space identifier, building function category, building structure type, building area, building height, construction year, and implicit carbon emissions over its entire life cycle.

[0110] Specifically, a building-embedded carbon spatial database can be constructed using individual buildings as the smallest spatial unit. This database includes a building spatial identifier to uniquely identify each building; the building function category is determined by combining point-of-interest data or urban census data; the building structure type is predicted by a graph neural network model; the building area is calculated from the building vector boundary; the building height is extracted from street view cloud data or digital surface models; the construction year is obtained by combining urban building census data or remote sensing image time-series analysis; and the total life-cycle embedded carbon emissions are predicted and output by a graph neural network model. This database stores the spatial mapping results in a structured manner, providing a unified data foundation for subsequent mapping and analysis.

[0111] S5.3: Generate a spatial distribution map of the implicit carbon throughout the entire life cycle of buildings at the regional scale; Based on attribute information from the building implicit carbon database, this embodiment generates regional-scale spatial distribution maps of building implicit carbon in both vectorized and rasterized forms. Vectorized mapping uses building boundary vectors as units, assigning the total life-cycle implicit carbon emissions to building polygons to achieve a refined representation of building-level implicit carbon distribution. Rasterized mapping, depending on research needs and spatial resolution requirements, converts the vectorized results into grid data of different spatial resolutions, achieving continuous spatial visualization through interpolation and weighted averaging, suitable for presenting macro-regional carbon emission patterns and model validation.

[0112] See Figure 4 The embodiment of this invention uses a certain urban agglomeration as an example to predict the implicit carbon emissions. Different shades of color in the figure represent different carbon emission intensity levels, which intuitively shows the spatial heterogeneity of implicit carbon in buildings within the study area.

[0113] S5.4: Perform aggregated statistics and spatial analysis based on functional zones, administrative boundaries, or geographical units.

[0114] Based on the results of building lifecycle occult carbon mapping, this embodiment utilizes spatial statistical tools from a geographic information system (such as ArcGIS spatial statistics modules) to aggregate and calculate the total occult carbon emissions and average emission intensity of buildings at different functional zones, administrative boundaries, and natural geographic unit scales. It analyzes the differences in the occult carbon structure and spatial distribution characteristics of buildings in different regions. Furthermore, spatial autocorrelation analysis can be employed: the Global Moran's I index is used to quantify the spatial clustering characteristics of occult carbon emissions within the overall region, while the Local Moran's I index is used to identify and visualize high-value clustering areas, low-value clustering areas, and anomalous distribution areas of occult carbon emissions within the city.

[0115] Through the above steps, quantitative assessment and spatial visualization of implicit carbon emissions throughout the entire life cycle, from individual buildings to regional scales, have been achieved, providing data support for refined urban carbon emission management. Spatial autocorrelation analysis can identify high- and low-value carbon emission clusters, revealing the spatial pattern and evolution of urban carbon emissions, and providing a scientific basis for low-carbon urban planning and carbon reduction policy formulation.

[0116] In specific implementation, the method proposed in the technical solution of this invention can be automatically executed by those skilled in the art using computer software technology. System devices for implementing the method, such as computer-readable storage media storing the corresponding computer program of the technical solution of this invention and computer equipment including the computer program running the corresponding computer program, should also be within the protection scope of this invention.

[0117] The electronic device for assessing building hidden carbon based on multimodal remote sensing and graph neural networks provided by the present invention will be described below. The electronic device for assessing building hidden carbon based on multimodal remote sensing and graph neural networks described below can be referred to in correspondence with the building hidden carbon assessment method based on multimodal remote sensing and graph neural networks described above.

[0118] The electronic device may include a processor, a communications interface, memory, and a communication bus. The processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions from the memory to execute a building-based implicit carbon assessment method using multimodal remote sensing and graph neural networks, primarily including the software processing components described above.

[0119] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0120] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the software processing portion of the building implicit carbon assessment method based on multimodal remote sensing and graph neural networks provided by the above methods.

[0121] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the software processing portion of the building implicit carbon assessment method based on multimodal remote sensing and graph neural networks provided by the above methods.

[0122] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0123] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0124] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for assessing building-hidden carbon based on multimodal remote sensing and graph neural networks, characterized in that, Includes the following steps: Construct a classification system for building structure types that address the implicit carbon footprint throughout the entire building lifecycle; Based on the building engineering list, identify building structure types and calculate implicit carbon throughout the entire life cycle to generate a sample dataset; Building feature extraction and fusion are performed based on multimodal remote sensing data to generate a unified building feature vector; Based on a multi-task graph neural network, the sample dataset and the building feature vector are jointly trained to obtain a model that can simultaneously predict building structure type and implicit carbon emissions throughout the entire life cycle. Based on the prediction results of the model, mapping and spatial analysis of the implicit carbon throughout the building's entire life cycle are performed.

2. The building-embedded carbon assessment method based on multimodal remote sensing and graph neural networks according to claim 1, characterized in that, The building structure type classification system includes a preset set of building structure types, a set of building life cycle stages, and a set of building materials. It also introduces energy technology indices to correct the carbon emission coefficients of materials for different regions.

3. The building-embedded carbon assessment method based on multimodal remote sensing and graph neural networks according to claim 1, characterized in that, The method of identifying building structure types and calculating implicit carbon throughout the entire life cycle based on the building project list specifically includes: Extract structural-related elements from the building project list to identify the building's structural type; The inventory data is categorized according to the preset life cycle stages, the implicit carbon emissions of each building material at each stage are calculated, and the total implicit carbon emissions of a single building throughout its entire life cycle are summarized to form a sample dataset.

4. The building-embedded carbon assessment method based on multimodal remote sensing and graph neural networks according to claim 1, characterized in that, The building feature extraction and fusion based on multimodal remote sensing data specifically includes: Acquire multi-source remote sensing data, including UAV oblique imagery, street view cloud data, and high-resolution satellite imagery; Registration and spatial alignment of multi-source data; Extracting architectural visual features from UAV imagery using a pre-trained visual model; The geometric and material features of facades in street view point clouds are extracted using a point cloud analysis model. Extracting spectral characteristics of building materials using high-resolution satellite imagery; The three features mentioned above are dynamically fused through an attention mechanism to generate a multimodal fusion feature vector.

5. The building-embedded carbon assessment method based on multimodal remote sensing and graph neural networks according to claim 1, characterized in that, The multi-task graph neural network uses individual buildings as nodes and constructs a building relationship graph based on edge weights determined by a combination of spatial distance between buildings, functional similarity, and differences in construction age. The node features are the multimodal fusion feature vectors; the network adopts a multi-layer graph sampling aggregation architecture, with each layer performing weighted sampling according to edge weights, and the number of samples gradually decreasing as the number of layers increases; the network shares parameters in the feature extraction layer, and the output layer is divided into two parallel branches, which are used for building structure type classification and full life cycle implicit carbon emission regression, respectively.

6. The building-embedded carbon assessment method based on multimodal remote sensing and graph neural networks according to claim 5, characterized in that, The training of the multi-task graph neural network adopts a joint loss function, which includes structural classification loss, implicit carbon emission regression loss, structural smoothing loss for constraining the consistency of prediction of adjacent building structures, and physical law loss for constraining the positive correlation between carbon emissions and volume of the same type of building.

7. The building-embedded carbon assessment method based on multimodal remote sensing and graph neural networks according to claim 1, characterized in that, The cartography and spatial analysis specifically include: The model prediction results are mapped to geospatial data to establish a spatial database of building-hidden carbon based on individual buildings. Generate a regional-scale spatial distribution map of the implicit carbon footprint throughout the entire life cycle of buildings; Based on functional zoning or administrative boundaries, aggregate statistics and spatial autocorrelation analysis are conducted to identify the spatial clustering characteristics and distribution patterns of carbon emissions.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the building implicit carbon assessment method based on multimodal remote sensing and graph neural networks as described in any one of claims 1 to 7.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the building implicit carbon assessment method based on multimodal remote sensing and graph neural networks as described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the building implicit carbon assessment method based on multimodal remote sensing and graph neural networks as described in any one of claims 1 to 7.

Images