Carbon emission downscaling method based on multi-source remote sensing images and building form large model

Abstract

The application discloses a carbon emission downscaling method based on multi-source remote sensing and a building form large model, and belongs to the technical field of carbon emission data processing. The method comprises the following steps: obtaining multi-source remote sensing images of a target region and performing preprocessing and fusion on the multi-source remote sensing images, inputting the fused images into a building form large model, outputting form feature vectors and geographical position coordinates of building instances, obtaining coarse-grained carbon emission grid data, constructing a form feature matrix according to a grid to which the building instances belong, inputting the form feature matrix into a carbon emission distribution network, outputting distribution weights, distributing total grid carbon emissions to each building instance according to the weights, and generating carbon emission spatial distribution data at a building scale. The application can be used for carbon emission monitoring at a building scale in a city, and improves the physical interpretability of carbon emission downscaling results.

Description

Technical Field

[0001] This invention relates to the field of carbon emission data processing technology. More specifically, this invention relates to a carbon emission downscaling method based on multi-source remote sensing imagery and large-scale building morphology models. Background Technology

[0002] Carbon emission monitoring is a crucial foundation for urban carbon management and emission reduction decisions. Currently, the acquisition of large-scale carbon emission data mainly relies on two approaches: one is a bottom-up accounting method based on fossil fuel consumption statistics, and the other is a top-down estimation method based on atmospheric observation and inversion models. Among these, gridded emission inventories based on energy statistics are widely used in the field of carbon emission spatialization research due to their advantage of explicit spatial representation. These data products spatially allocate the total regional carbon emissions according to proxy variables such as population distribution, nighttime light intensity, and GDP, generating carbon emission distribution maps in the form of regular grid cells, with grid cell side lengths typically ranging from 500 to 1000 meters.

[0003] The generation process of the aforementioned carbon emission grid data itself includes the meaning of spatial downscaling. However, the spatial resolution of its final product is still insufficient to support the needs of emission source tracing and refined management at the building scale. To achieve further downscaling from coarse-grained grids to individual buildings, existing technologies mainly continue the idea of ​​weighting proxy variables. The proxy variables used include building volume, building density, land use type, etc. When dealing with carbon emission allocation at the building scale, the following problems are faced: there is only a statistical correlation between proxy variables and actual building energy consumption, and there is a lack of causal relationship at the physical mechanism level. Two buildings with similar volumes may have a difference of several times in actual carbon emissions due to differences in the performance of the building envelope, usage function, and operation mode. Existing proxy variable weighting methods cannot distinguish such differences.

[0004] Multi-source remote sensing images contain detailed geometric information such as building height, base area, and roof shape, which theoretically can serve as a more physically interpretable basis for downscaling. However, in practical applications, extracting building morphology parameters from multi-source remote sensing images faces several technical challenges. Images acquired by different sensors differ in imaging mechanisms, spatial resolution, and radiometric characteristics. High-precision registration and effective fusion of these images, ensuring that the fused image fully preserves various complementary information, are the technical prerequisites for subsequent building morphology parameter extraction. At the same time, the spatial variation in building density in urban built environments is significant, with the number of building instances within a coarse-grained grid cell ranging from a few to hundreds. How to design an allocation model that can adapt to this variable input scale is a problem that needs to be solved in the engineering application of downscaling methods.

[0005] Furthermore, there are very few directly quantifiable and publicly available datasets for carbon emissions at the building scale. When training carbon emission allocation models, the supervision signals can only come from the total carbon emissions at the grid level. Under these conditions of lacking ground truth labels for building-by-building emissions, how to guide the model to learn an allocation that conforms to the physical laws of building energy consumption is a technical problem that needs to be solved. Summary of the Invention

[0006] This invention provides a carbon emission downscaling method based on multi-source remote sensing imagery and large-scale building morphology models, which can be used for carbon emission monitoring at the urban building scale and improves the physical interpretability of carbon emission downscaling results.

[0007] To achieve these objectives and other advantages according to the present invention, a carbon emission downscaling method based on multi-source remote sensing imagery and a large-scale building morphology model is provided, comprising the following steps: Acquire multi-source remote sensing images of the target area, including optical remote sensing images with a spatial resolution of 0.5 to 1 m and synthetic aperture radar images with a spatial resolution of 1 to 3 m. Perform radiometric calibration and atmospheric correction on the optical remote sensing images, perform multi-view processing on the synthetic aperture radar images, and register and fuse the processed images to obtain fused image data. The fused image data is input into a pre-trained large-scale building morphology model. The fused image data is segmented into building instances, and the building outline of each building instance and the morphological feature vector composed of building height, building base area and building shape coefficient are output. The coordinates of the center point of the bounding rectangle of each building outline are calculated as the geographical coordinates of each building instance. The large-scale building morphology model is a deep neural network trained with sample multi-source remote sensing images and corresponding labeled data of building height, building base area and building shape coefficient. Obtain coarse-grained carbon emission grid data for the target area. The grid cell side length of the coarse-grained carbon emission grid data is within 500~1000m, and each grid records a total carbon emission value. Based on the geographical coordinates of each building instance, all building instances whose center point coordinates fall into the same grid are grouped into the building instance set of that grid. For each grid, the morphological feature vectors of the N building instances included in the grid are concatenated row by row to construct a building morphological feature matrix of dimension N×3. The building morphological feature matrix is ​​input into a pre-trained carbon emission allocation network, which outputs the carbon emission allocation weight of each building instance in the grid, and satisfies that the sum of the carbon emission allocation weights of all building instances in the grid is 1. The carbon emission allocation network is a neural network. For each grid, its total carbon emissions are allocated according to the carbon emission allocation weights of each building instance within that grid, generating spatial distribution data of carbon emissions at the building scale.

[0008] Preferably, the carbon emission allocation network is a neural network based on a deep set architecture. The deep set architecture includes a shared encoder network and a permutation-invariant aggregation function. The encoder network independently maps each row of the morphological feature vector in the N×3 building morphological feature matrix to a high-dimensional latent space. The aggregation function aggregates all mapped latent vectors into a fixed-length global representation vector. Then, based on the global representation vector, the carbon emission allocation weight of each building instance is decoded and output to adapt to the change in the number N of building instances in the grid.

[0009] Preferably, for a building instance whose building outline covers two or more coarse-grained carbon emission grids, the carbon emission allocation weight of the building instance is weighted by using the proportion of the area of ​​the intersection between the building instance and each grid to the area of ​​its building base as the area weight. The carbon emission allocation weight of the building instance is then weighted and corrected by using the area weight as a coefficient. After correction, the sum of the allocation weights of all building instances in each grid is 1, and then they participate in the allocation of the total carbon emission value in their respective grids.

[0010] Preferably, the processed images are registered and fused to obtain fused image data, specifically including: Building outlines are pre-extracted from the processed optical remote sensing images to generate binary optical building outline templates; Using the optical building outline template as a reference, the processed synthetic aperture radar image is matched and registered by outline feature matching. The RANSAC algorithm is used to remove mismatched point pairs so that the final registration error does not exceed 0.3 optical image pixels. The SwinFusion algorithm based on cross-modal attention mechanism is used to perform feature-level fusion of the registered images. In this algorithm, optical remote sensing image features are used as queries and synthetic aperture radar (SAR) image features are used as keys and values. Cross-modal feature interaction is achieved through cross-attention calculation. During the fusion process, an identity mapping is applied to the phase components of the SAR image features to generate fused image data.

[0011] Preferably, the large building form model includes a three-dimensional reconstruction sub-network. The three-dimensional reconstruction sub-network takes fused image data as input, firstly generates a disparity map by the monocular depth estimation module and converts it into a normalized digital surface model, and then outputs the voxel occupancy probability of the building instance by the voxel reconstruction module. Based on the voxel occupancy probability, the ratio of the external surface area to the volume is calculated to obtain the building form coefficient value.

[0012] Preferably, the large building form model uses SwinTransformer as the backbone network, with three parallel task heads set at the end: building instance segmentation head, building height regression head, and building shape coefficient regression head. The training loss function for the large-scale architectural model adopts a dynamically weighted multi-task loss function, and the calculation formula is as follows: L total = α(t)×L seg +β(t)×L height +γ(t)×L shape , Among them, L seg L is a weighted sum of the Dice loss and cross-entropy loss for the segmentation task. height For Huber loss in high regression tasks, L shape For the Huber loss in the body size coefficient regression task, α(t), β(t), and γ(t) are dynamic weights that are adaptively adjusted with the training epoch t. The adjustment rule for the dynamic weights is as follows: After every K training rounds, calculate the current stage L. seg L height L shape The respective rate of decline is calculated as the ratio of the reciprocal of the rate of decline of each loss in the current stage to the sum of the reciprocals of the three losses, and after being corrected by the smoothing coefficient τ, it is used as the dynamic weight of each loss in the next stage, where K takes the value of 5~10 and τ takes the value of 0.8~0.95.

[0013] Preferably, before constructing the N×3 building morphology feature matrix, the following steps are also included: The architectural function type feature values ​​of each building instance are extracted from the fused image data using a large architectural morphology model. The building function type feature value is concatenated with the original form feature vector to generate an extended form feature vector with a dimension of 4. For each grid, the extended morphological feature vectors of the N building instances included in the grid are concatenated row by row to construct a building morphological feature matrix with a dimension of N×4. In subsequent steps, an N×4 building morphology feature matrix is ​​input into a pre-trained carbon emission allocation network.

[0014] Preferably, the carbon emission allocation network is trained using a weakly supervised learning framework, and the loss function consists of two parts: a total constraint term and a comparison constraint term. L = L mse +λ×L contrast ; Among them, L mse For the total carbon emissions constraint term in the grid, L is calculated as the mean square error between the actual total carbon emissions of the grid and the sum of the carbon emissions allocated to all buildings within that grid. contrast To compare the constraint terms, the variance of carbon emission intensity of buildings of the same functional type within the same grid is calculated as the ratio of the variance of carbon emission intensity of buildings of different functional types. λ is the balance coefficient, with a value ranging from 0.1 to 0.5.

[0015] Preferably, for each grid, its total carbon emissions are allocated according to the carbon emissions allocation weights of each building instance within that grid, preceded by the following steps: Obtain land use classification data for the target area, which includes building land, industrial land, transportation land, green space, and water area; For each grid, calculate the proportion R of the built-up area to the total area of ​​the grid based on the land use classification data. build ; Multiply the total carbon emissions of this grid by R build This yields the total building carbon emissions for that grid. In subsequent steps, for each grid, its total building carbon emissions value is allocated according to the carbon emission allocation weight of each building instance within that grid.

[0016] Preferably, the carbon emission allocation network employs Monte Carlo dropout during training and inference, adding a Dropout layer with a dropout rate of 0.1 to 0.2 after each fully connected layer; By performing 20 to 50 forward propagations on the building morphology feature matrix of the same grid, multiple sets of carbon emission allocation weights are obtained. The mean and standard deviation of carbon emissions allocated to each building instance corresponding to each weight group are calculated to generate spatial distribution data with uncertainty in carbon emissions per building.

[0017] The present invention has at least the following beneficial effects: First, this invention extracts building height, base area, and shape coefficient values ​​from multi-source remote sensing images using a large-scale building morphology model, constructs a physically meaningful morphological feature vector, replaces traditional socio-economic proxy variables, and achieves scientific allocation for each building under grid-based total carbon emission constraints, generating spatial distribution data of building-scale carbon emissions.

[0018] Secondly, this invention solves the problem that fixed-dimensional networks cannot adapt to the differences in the number of buildings in different grids by using a carbon emission allocation network with a deep set architecture, and by using a shared encoder and a permutation-invariant aggregation function to process a variable number of building instances. It does not require zero-padding or truncation of the input matrix.

[0019] Third, this invention uses the SwinFusion algorithm with cross-modal attention mechanism to perform feature-level fusion of optical and synthetic aperture radar images and applies an identity mapping to the SAR phase components. While preserving optical texture information, it also fully preserves phase information, providing a high-quality multimodal data foundation for building height and 3D morphology inversion.

[0020] Fourth, this invention introduces a contrast constraint term into the weakly supervised loss function to minimize the variance of carbon emission intensity of buildings with the same functional type and maximize the variance of carbon emission intensity of buildings with different types, thereby guiding the network to learn an allocation strategy that conforms to physical laws and suppressing the spurious correlations that may be generated by training with pure total constraint.

[0021] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Detailed Implementation

[0022] The present invention will now be described in further detail to enable those skilled in the art to implement it based on the description. It should be understood that terms such as "having," "comprising," and "including" as used herein do not exclude the presence or addition of one or more other elements or combinations thereof. It should be noted that the experimental methods described in the following embodiments, unless otherwise specified, are conventional methods and therefore should not be construed as limiting the present invention.

[0023] A carbon emission downscaling method based on multi-source remote sensing imagery and large-scale building morphology models includes the following steps: The target area is the region where building-scale carbon emission accounting is required. Multi-source remote sensing imagery of the target area is acquired, including optical remote sensing images with a spatial resolution of 0.5–1 m and synthetic aperture radar (SAR) images with a spatial resolution of 1–3 m. Optical remote sensing images can clearly distinguish the outlines of individual buildings and roof details, while SAR images include both amplitude and phase information. Radiometric calibration and atmospheric correction are performed on the optical remote sensing images. The digital count values ​​of the optical remote sensing images are converted into apparent radiance values, and then the apparent radiance values ​​are converted into surface reflectance values. Multi-look processing is performed on the SAR images, with the number of looks determined based on the original image resolution and the target resolution. While suppressing speckle noise, acceptable spatial resolution is maintained. The processed images are registered and fused. During the registration process, optical remote sensing imagery is used as the reference image and synthetic aperture radar imagery is used as the image to be registered. Feature points such as building corners and road intersections are selected as control points. A polynomial transformation model is used to establish the geometric mapping relationship between the two images. The synthetic aperture radar imagery is resampled to the same spatial reference and pixel size as the optical remote sensing imagery using bilinear interpolation. The registration accuracy is controlled within 0.5 pixels. During the fusion process, spectral texture features are extracted from the optical remote sensing imagery, and structural and texture features are extracted from the synthetic aperture radar imagery. The two types of features are combined at the feature layer to obtain fused image data, which has both the spectral texture information of the optical imagery and the geometric structure information of the synthetic aperture radar imagery. The fused image data of the target area is input into a pre-trained large-scale building morphology model. The fused image data is processed pixel-by-pixel to segment building instances, outputting a mask region for each building instance. Vector boundaries of the building contours are extracted from these mask regions, outputting the building contours of each instance and a morphological feature vector composed of building height, building base area, and building shape coefficient values. The building height is the vertical distance from the ground to the top of the roof; the building base area is the horizontal projected area enclosed by the building contour; and the building shape coefficient is the ratio of the building's external surface area in contact with the outdoor atmosphere to the volume it encloses. The coordinates of the center point of the circumscribed rectangle of each building contour are calculated as the geographical coordinates of each building instance. Specifically, for each building contour vector polygon, its circumscribed rectangle is calculated, and the coordinates of the geometric center point of the circumscribed rectangle are taken as the geographical coordinates of that building instance. The geographical coordinates use the same projection coordinate system as the subsequent coarse-grained carbon emission grid data. The large-scale building morphology model is a deep neural network trained using sample multi-source remote sensing images and corresponding labeled data of building height, building base area, and building shape coefficient. The model uses a large number of multi-source remote sensing images as input, with each image corresponding to a set of labeled data. The labeled data includes the building height, building base area, and building shape coefficient of each building instance within the image range. The building height can be labeled using LiDAR point cloud data or stereo image pair matching. The building base area can be labeled using existing GIS building outline data or manual vectorization of local areas. The building shape coefficient can be labeled using building information modeling or architectural design drawings. In practical applications, the model is first pre-trained using publicly available remote sensing images and building height datasets. Then, the model is fine-tuned using a small number of precisely labeled target area samples. The number of building instances in the fine-tuning samples can be in the hundreds to thousands. During model training, a supervised learning approach is used. The sample images are input into the building morphology model, which outputs the predicted building outline, height, base area, and shape coefficient. The predicted values ​​are compared with the labeled values, the loss function is calculated, and the model parameters are updated using the backpropagation algorithm. After sufficient training, the building morphology model has the ability to automatically extract building instances and their morphological parameters from fused image data. To obtain coarse-grained carbon emission grid data for the target area, the target area is divided into regular grid units. Each grid unit records a spatial dataset of total carbon emissions, which is the coarse-grained carbon emission grid data. The grid unit side length of the coarse-grained carbon emission grid data is within 500~1000m. Each grid records a total carbon emission value. The statistical scope of the total carbon emission value covers both direct and indirect carbon emissions during the building operation phase. It is consistent with the accounting boundary of the building-scale carbon emission data to be generated later. When obtaining this data, it is necessary to confirm its spatial reference coordinate system and time base to ensure that it adopts a unified coordinate system and a similar time range with the geographical coordinates of the building instance. Based on the geographical coordinates of each building instance, all building instances are traversed. For each building instance, the grid cell containing its center point coordinates is determined. All building instances whose center point coordinates fall within the same grid are grouped into the building instance set of that grid. After traversal, each grid obtains a list of building instances. For grids with zero building instances, no further processing is performed. For each non-empty grid, the number N of building instances in its building instance set is counted. The morphological feature vectors of the N building instances included in that grid are concatenated row-wise to construct a building morphological feature matrix of dimension N×3. Each row of the matrix corresponds to a building instance. The first column stores the building height value, the second column stores the building base area value, and the third column stores the building shape coefficient value. The building morphological feature matrix is ​​input into a pre-trained carbon emission allocation network. The input dimension of the network is the same as that of the building morphology. The feature matrix is ​​dimensionally adapted, and the network's output layer outputs N real values. After processing with a normalized exponential function, N positive values ​​summing to 1 are obtained, which serve as the carbon emission allocation weights for each building instance within the grid. The output of each building instance's carbon emission allocation weight satisfies the condition that the sum of all building instance carbon emission allocation weights within the grid is 1. The magnitude of the carbon emission allocation weight represents the relative share of each building instance in the total carbon emissions of the grid. The carbon emission allocation network is a neural network, preferably a fully connected network or a convolutional network. The training data is obtained by collecting coarse-grained carbon emission grid data and corresponding building morphology feature matrix data for multiple cities, covering different built environment density types, including high-density central urban areas, medium-density general urban areas, and low-density suburbs. Each city's area is selected from 10 to 50 square kilometers. 2 The target area was analyzed by extracting building instances one by one according to grid cells. A building morphology feature matrix was constructed as the training input sample, and the total carbon emission value of the grid cells was used as the supervision label. The dataset was divided into training, validation, and test sets in an 8:1:1 ratio. The Adam optimizer was used during training, with an initial learning rate set to 1×10⁻⁶. -4 ~to 1×10 -3The batch size is 16~64. The loss function is based on the mean square error between the sum of carbon emissions of each building after allocation and the total amount of grid annotation. The network parameters are updated through the backpropagation algorithm until the validation set loss converges, so that the network learns the mapping relationship between building morphological features and carbon emission allocation weights. For each grid, its total carbon emissions are allocated according to the carbon emission weights of each building instance within that grid, generating spatial distribution data of carbon emissions at the building scale. The total carbon emissions for that grid are E. sub, total The carbon emission allocation weight of the i-th building instance within the grid is w. i Then the carbon emissions E allocated to this building instance sub, i Calculate using the following formula: E sub, i = E sub, total × w i , Among them, E sub, total w represents the total carbon emissions recorded by this grid. i Assign a weight to the carbon emissions of the i-th building instance within the grid, where i ranges from 1 to N, and N is the total number of building instances within the grid. sub, i The carbon emissions allocated to the i-th building instance; By traversing all non-empty grids and merging the allocation results of each grid, the carbon emission estimate of each building instance within the target area is obtained. These carbon emission values ​​are then associated with the building outline vector data to generate spatial distribution data of carbon emissions at the building scale. This data uses the building outline vector surface as the spatial unit, and the attribute table of each surface feature records the annual carbon emission value of the building instance, in tons of carbon dioxide equivalent per year.

[0024] In the above technical solution, building height, building base area, and building shape coefficient values ​​are extracted from multi-source remote sensing images through a large building morphology model, and a physically meaningful building morphology feature vector is constructed. Using this feature vector as input, the mapping relationship between building morphology and carbon emissions is learned through a carbon emission allocation network. Under the constraint of grid-based total carbon emissions, building-scale carbon emission spatial distribution data is generated, which significantly improves the resolution and physical interpretability of the carbon emission spatialization results.

[0025] The carbon emission allocation network receives the building morphology feature matrix of each grid as input. The number N of building instances included in each coarse-grained carbon emission grid varies with the density of the urban built environment, ranging from a few to hundreds. Ordinary feedforward neural networks require a fixed input dimension and cannot directly input variable-sized building morphology feature matrices into the network for processing. In another technical solution, the carbon emission allocation network is a neural network based on a deep set architecture. The deep set architecture includes a shared encoder network and a permutation-invariant aggregation function. The encoder network independently maps each row of the N×3 building morphology feature matrix, i.e., the three-dimensional morphology feature vector of a single building instance, to a high-dimensional latent space. For the j-th building instance, let its input feature vector be x. j The dimension is 3, including building height, building base area, and building shape coefficient. The encoder network consists of several fully connected layers and activation functions, for x j Perform a layer-by-layer transformation to output a latent vector h of dimension d. j The transformation process is represented as follows: h j = φ( x j ) Where φ represents the mapping function represented by the encoder network, d is the dimension of the latent space, and the value is between 64 and 256. The same encoder network shares parameters among all N building instances in the grid. The feature mapping rules of each building instance are completely consistent, and the mapping process does not depend on the arrangement position of the building instance in the matrix. After processing N building instances, the encoder network obtains N latent vectors of dimension d. An aggregation function aggregates all the mapped N latent vectors into a fixed-length global representation vector, denoted as g. The aggregation operation uses a summation function to satisfy permutation invariance while preserving the information about the number of building instances, enabling the subsequent decoder to perceive the overall scale of the buildings within the grid. This is represented as: Then, based on the global representation vector, the carbon emission allocation weights of each building instance are decoded and output to adapt to the change in the number N of building instances in the grid. The aggregated global representation vector g is input into the decoder network, which also consists of several fully connected layers. Its output dimension is N, corresponding to N building instances in the grid. Let the original value of the decoder output be s. j If the value of j ranges from 1 to N, then the carbon emission allocation weight w for each building instance is... j Calculated using the normalized exponential function: The original predicted value for the k-th building instance is s. k The value of k ranges from 1 to N; The output layer of the carbon emission allocation network uses the Softmax activation function to normalize the original predicted value of each building instance output by the decoder, so that the carbon emission allocation weight of all building instances is between 0 and 1, and the sum of the allocation weights of all building instances in the same grid is 1.

[0026] In the above technical solution, a deep set architecture is used as the carbon emission allocation network. The morphological feature vector of each building instance is independently mapped to a high-dimensional latent space through a shared encoder network. Then, the permutation-invariant aggregation function is used to aggregate all latent vectors into a global representation vector of fixed length. This enables the network to process any number of building instances N without zero-padding or truncation of the input, thus improving the model's adaptability to grids with different building densities and the allocation accuracy.

[0027] The grid to which a building falls is determined by the coordinates of the center point of the bounding rectangle of the building outline. When a building outline crosses the boundaries of more than two coarse-grained carbon emission grids, simply determining its grid affiliation based on the coordinates of the building's center point would lead to the building's overall carbon emissions being incorrectly allocated to a single grid, resulting in discontinuous carbon emission allocation at the grid boundaries. In another technical solution, after the initial calculation of the carbon emission allocation weights for each building instance is completed, a boundary crossing check is performed on each building instance. The vector polygon of the building outline is spatially overlaid with the grid surface of the coarse-grained carbon emission grid, and the number of grid cells with which the building outline has spatial intersection relationships is counted. If the number of intersecting grids is 1, it means that the building is completely located inside a single grid and no correction is needed. The original allocation weights output by the carbon emission allocation network are directly used. If the number of intersecting grids is greater than or equal to 2, it means that the building is a cross-grid building. For building instances whose outlines cover two or more coarse-grained carbon emission grids, the carbon emission allocation weight of the building instance is weighted and corrected by using the proportion of the area of ​​the intersection between the building instance and each grid to its building footprint area as the area weight. The original allocation weight is then split according to the area ratio, so that the building receives an allocation weight component proportional to its footprint in each intersecting grid. Specifically, for the m-th grid intersecting with the building instance, the corrected allocation weight w of the building instance within that grid is... sub, m Calculate using the following formula: w sub, m = wo rig × r m Among them, wo rig r represents the original allocation weights output by the carbon emission allocation network for this building instance. m This is the area weight of the building instance in the m-th intersecting grid. For grids where the building does not intersect, its weight is 0. After completing the weight correction for all cross-grid buildings, each grid needs to be re-normalized because the weights of cross-grid buildings have been split across multiple grids, and these grids may also receive weight components from other cross-grid buildings. The sum of the allocated weights within each grid may no longer equal 1. For each grid, the current allocated weights of all building instances within that grid are summed to obtain the total weight S. Then, the allocated weight of each building instance within that grid is divided by S to obtain the final normalized allocated weight. After correction, the sum of the allocated weights of all building instances within each grid is 1, and then they participate in the allocation of the total carbon emission value within their respective grids.

[0028] In the above technical solution, the carbon emission allocation weight is adjusted by calculating the ratio of the spatial intersection area of ​​the building instance and each intersecting grid to the building base area. The carbon emissions of the building are allocated to each intersecting grid according to its actual spatial distribution ratio. After the adjustment, the sum of the allocation weights in each grid still meets the normalization condition, eliminating the boundary effect and allocation error caused by the building crossing the grid boundary.

[0029] Traditional registration methods are prone to producing mismatched point pairs in densely built-up areas, and traditional fusion algorithms are prone to losing or distorting phase information when processing synthetic aperture radar (SAR) images, affecting the accuracy of subsequent building height and 3D shape inversion. In another technical solution, the processed images are registered and fused to obtain fused image data, specifically including: The processed optical remote sensing image is a surface reflectance image that has undergone radiometric calibration and atmospheric correction. Building outlines are pre-extracted from the processed optical remote sensing image. The Canny edge detection operator is used to process the optical remote sensing image to extract pixels with drastic grayscale changes, forming an edge intensity map. The edge intensity map is binarized, and an empirical threshold is set. Pixels with edge intensities higher than the threshold are assigned a value of 1, and the rest are assigned a value of 0, resulting in an initial binary edge map. Morphological closing operations (dilation followed by erosion) are performed on the initial binary edge map to connect broken edge segments and fill small holes. Connected regions are extracted, and small connected regions with areas significantly smaller than the typical building outline area are removed. Connected regions corresponding to the building outline are retained, generating a binary optical building outline template. The pixel value of the building outline area is 1, and the pixel value of the background area is 0. The processed synthetic aperture radar (SAR) image is an intensity image that has undergone multi-look processing. Using the generated optical building outline template as a reference, contour feature matching and registration are performed on the processed SAR image. Using the spatial reference of the optical remote sensing image as a reference, the SAR image is registered to the optical image. Nodes and corner points of the building outlines are extracted from the optical building outline template as registration control points. These points are located at significant turning points on the building edges and have good spatial positioning. Simultaneously, the same contour extraction process is used to extract building outline features from the SAR intensity image, obtaining corresponding feature points in the SAR image. An initial matching relationship is established between the two sets of feature points. Using a feature descriptor-based method, the gradient direction histogram of the neighborhood of each feature point is calculated as a descriptive vector. Nearest neighbor matching is performed using the Euclidean distance between the descriptive vectors to obtain an initial set of matched point pairs. Because densely built areas contain a large number of buildings with similar shapes and regular arrangements... The initial set of matching point pairs inevitably includes mismatched point pairs. The RANSAC algorithm is used to remove mismatched point pairs. The minimum number of matching point pairs required to solve the transformation model are randomly selected from the initial set of matching point pairs. For the polynomial transformation model, the parameters of a candidate transformation model are estimated using these point pairs. The candidate model is used to test all other matching point pairs. The position deviation of each point pair after model transformation is calculated. Point pairs with position deviation less than a preset threshold are marked as inliers, and those with deviation greater than the threshold are marked as outliers. The number of inliers corresponding to the candidate model is recorded. The above random sampling and model verification process is repeated. After the iteration, the candidate model corresponding to the iteration with the most inliers is selected as the final transformation model. The corresponding set of inliers is the correct matching point pairs after removing mismatches. The SAR image is resampled to the pixel grid of the optical image using the final transformation model, so that the final registration error does not exceed 0.3 optical image pixels. The SwinFusion algorithm, based on a cross-modal attention mechanism, is employed to perform feature-level fusion on the registered images. Optical remote sensing image features are used as queries, and synthetic aperture radar image features are used as keys and values. Cross-modal feature interaction is achieved through cross-attention computation. The optical feature map is F. opt SAR feature map is F sar For F respectively opt and F sar Perform linear projection to generate a query matrix Q, a key matrix K, and a value matrix V, where Q is derived from F. opt Through the weight matrix W q K is obtained by projection, and F is obtained by projection. sar V is obtained by projecting the weight matrix Wk, and V is derived from F. sar Through the weight matrix W vProjection yields that, for each location in the optical feature map, the similarity between its query vector and the key vectors of all SAR feature locations is calculated. After scaling and normalization, attention weights are obtained. These weights are then used to perform a weighted summation of the SAR feature value vectors. Cross-modal enhancement features are obtained by actively querying SAR features from optical features. The calculation process of cross-modal attention is expressed as follows: Among them, K T d represents the transpose of the bond matrix K. k The dimension of the key vector is used to scale the dot product result to prevent gradient vanishing. Through this cross-attention computation, optical features and SAR features achieve cross-modal information interaction at the feature layer, and optical features selectively extract complementary structural and texture information from SAR features; In the feature extraction stage, the phase information of SAR complex images is organized into independent phase feature channels. In cross-modal attention computation, the phase feature channels do not participate in the linear projection transformation of Q, K, and V. Instead, they are directly transmitted to the decoder input through identity connections in the fusion layer. During the fusion process, identity mapping is applied to the phase components of the synthetic aperture radar image features, and the output phase features are equal to the input phase features. The phase information of the SAR image remains intact during the fusion process. The feature map after cross-modal attention fusion generates fused image data through the decoder. The fused image has both the spectral texture expression capability of optical images and the geometric structure information of SAR images, and retains the complete phase channels.

[0030] In the above technical solution, a binary template is generated by pre-extracting the building outline as a registration reference. Combined with the RANSAC algorithm to remove mismatched point pairs, the registration error is controlled within 0.3 optical image pixels. The SwinFusion algorithm with cross-modal attention mechanism is adopted to perform cross-attention calculation with optical features as queries and SAR features as keys and values. An identity mapping is applied to the SAR phase components to retain the phase information while preserving the optical texture information, thus providing high-quality fusion data for subsequent large-scale building morphology models.

[0031] The building form factor is an important parameter reflecting the thermal performance of a building. Its definition involves the ratio of the building's external surface area to its volume. It is difficult to directly obtain accurate three-dimensional external surface area and volume data from a single two-dimensional remote sensing image. In another technical solution, the large building form model includes a three-dimensional reconstruction sub-network. This sub-network takes fused image data as input. First, a monocular depth estimation module performs pixel-by-pixel depth estimation on the fused image, generating a disparity map and converting it into a normalized digital surface model. The value of each pixel represents the vertical distance of the feature above the ground at that location; building areas have positive values, while ground areas approach zero. The normalized digital surface model is then input into a voxel reconstruction module, which converts the two-dimensional height field into a three-dimensional voxel mesh representation. Finally, the voxel reconstruction module outputs... The voxel occupancy probability of building instances addresses the occlusion and shadow issues present in densely built-up areas. During training, the monocular depth estimation module incorporates samples from occluded and shadowed scenes into the training data, enabling the network to learn the ability to infer the complete height and shape of occluded buildings from locally visible parts. For areas where depth estimation is uncertain due to shadows, the voxel reconstruction module outputs an occupancy probability close to 0.5, indicating uncertainty regarding the building boundaries in that area. In subsequent calculations of surface area and volume, the contribution of such areas is included in a probability-weighted manner to avoid bias introduced by rigid binary judgment. Using the building instance segmentation results, a voxel set belonging to each building instance is extracted from the scene voxels. For the voxel set of a building instance, let the occupancy probability of the q-th voxel be p. q The value ranges from 0 to 1. The volume V of this building example is calculated using the following formula: , Where Ω is the set of voxels of the building instance, and v is the volume of a single voxel; The outer surface area A is calculated by counting the faces in contact with air in the voxel set. For each voxel in the set, its six adjacent directions are traversed. If a neighboring voxel is not in the set or its occupancy probability is below a preset threshold, then the face in that direction is included in the outer surface area. Let the area of ​​that face be 'a', where 'a' is the area of ​​a single voxel face. The outer surface area A is then calculated using the following formula: , Where d is the adjacency direction index, p neighbor(q,d) Let f be the probability of the qth voxel being occupied by its neighboring voxels in the dth direction, and f be the decision function. f takes the value 1 when the neighboring voxel is a non-building voxel, and 0 otherwise. The ratio of external surface area to volume is calculated based on the voxel occupancy probability to obtain the building form factor value. The building form factor value S is calculated using the following formula: S = A / V Where A is the external surface area, V is the volume, and S is the building shape coefficient, which reflects the heat dissipation surface area corresponding to the unit volume of the building.

[0032] In the above technical solution, a three-dimensional reconstruction sub-network is set up inside the large model of the building form. The normalized digital surface model is recovered from the fused image data through the monocular depth estimation module. Then, the voxel reconstruction module outputs the three-dimensional voxel occupancy probability of the building instance. The ratio of the external surface area to the volume is directly calculated from the voxel geometric representation, which avoids the error caused by the simplified geometric assumption, improves the accuracy of the shape coefficient extraction, and enables carbon emission allocation to more accurately reflect the actual thermal performance of the building.

[0033] The large-scale architectural model needs to simultaneously perform three tasks: building instance segmentation, building height regression, and building shape coefficient regression. Segmentation is a pixel-by-pixel classification task, while height and shape coefficient regression are continuous value regression tasks. The loss functions for these different tasks differ in magnitude and convergence speed. When using a multi-task loss function with fixed weights, tasks with larger loss values ​​or steeper gradients tend to dominate the training process, leading to underfitting of other tasks. In another technical solution, the large-scale architectural model uses the SwinTransformer as the backbone network to extract multi-scale feature maps from the input fused image data. Shallow feature maps have higher resolution and include spatial detail information, while deep feature maps... To address the issue of low resolution and encoding semantic context information, three parallel task heads are set at the end of the backbone network: a building instance segmentation head, a building height regression head, and a building shape coefficient regression head. The building instance segmentation head receives multi-scale feature maps from the backbone network, fuses features of different scales through a feature pyramid structure, and after upsampling to restore the resolution to the input image, outputs the category of building instances pixel by pixel. The building height regression head receives deep feature maps from the backbone network, compresses the spatial dimension through global average pooling, and outputs the height prediction value of each building instance through a fully connected layer. The building shape coefficient regression head shares deep features with the height regression head and also outputs the shape coefficient prediction value through global average pooling and a fully connected layer. The training loss function for the large-scale architectural model adopts a dynamically weighted multi-task loss function, and the calculation formula is as follows: L total = α(t)×L seg +β(t)×L height +γ(t)×L shape , Among them, L seg For segmentation tasks, this is a weighted sum of Dice loss and cross-entropy loss. Dice loss measures the overlap between the predicted segmented region and the labeled region and is insensitive to class imbalance. Cross-entropy loss calculates the deviation between the predicted class probability and the true label pixel by pixel. The combination of the two takes into account both region overlap accuracy and pixel classification accuracy. heightFor high-resolution regression tasks, the Huber loss degenerates into mean absolute error when prediction bias is large, suppressing the amplification effect of outliers on the gradient. When prediction bias is small, it degenerates into mean squared error to ensure smooth convergence. shape For the Huber loss in the body size coefficient regression task, α(t), β(t), and γ(t) are dynamic weights that are adaptively adjusted with the training epoch t. The adjustment rule for the dynamic weights is as follows: After every K training rounds, calculate the current stage L. seg L height L shape The respective descent rate is calculated as the ratio of the reciprocal of the descent rate of each loss in the current stage to the sum of the reciprocals of the three losses. After being corrected by a smoothing coefficient τ, it serves as the dynamic weight of each loss in the next stage. The slower the descent rate of a task, the larger its candidate weight is, so as to obtain more training resources. The value of K is 5~10. To avoid the weights changing drastically due to fluctuations in a single stage, a smoothing coefficient τ is set. The value of τ is 0.8~0.95. The closer it is to 1, the more conservative the weight adjustment and the smoother the change. The sum of α(t), β(t), and γ(t) is always kept at 1. In the initial stage of training, α is set to 0.5, and β and γ are each set to 0.25, so that the segmentation task has a higher weight in the early stage and prioritizes the establishment of accurate building contour extraction capabilities. As training progresses, the descent rate of the segmentation task naturally slows down after convergence, and its weight is automatically reduced. The weights of the height regression and body shape coefficient regression tasks are increased accordingly, and training resources are automatically tilted towards the regression tasks that have not yet converged.

[0034] In the above technical solution, a dynamic weighted multi-task loss function is adopted. After every K training rounds, the dynamic weights are adaptively adjusted according to the rate of decrease of the loss value of each task. In the early stage of training, the segmentation task has a larger weight and prioritizes learning the building outline features to provide accurate building regions for subsequent regression tasks. In the later stage of training, the weights of height regression and body shape coefficient regression are gradually increased to optimize regression accuracy and balance the learning progress of each task.

[0035] Two buildings with similar shapes in a city may have energy consumption differences of several times due to different functions. If the carbon emission allocation network is based solely on three-dimensional geometric features, it is difficult to provide reasonable allocation results in areas with mixed functions. In another technical solution, before constructing a building shape feature matrix with dimensions N×3, the following steps are also included: The architectural morphology model extracts the architectural function type feature values ​​of each building instance from the fused image data. During the training phase, the architectural morphology model learns the ability to identify building function types from images. Specifically, based on the visual cues shown by buildings in remote sensing images, including roof material and color, geometric regularity of building outlines, spatial relationship between buildings and surrounding roads and green spaces, and building group arrangement patterns, the functional type classification system is set according to actual application needs, such as being divided into categories such as residential buildings, commercial buildings, office buildings, industrial buildings, and public buildings. During the model inference phase, the architectural morphology model outputs a functional type label for each building instance and converts the label into a scalar form of architectural function type feature value. The building function type feature value is concatenated with the original form feature vector. The original form feature vector has a dimension of 3, including the building height value, the building base area value, and the building shape coefficient value. The building function type feature value is added as a new scalar dimension. After being concatenated with the original vector, an extended form feature vector with a dimension of 4 is generated, which includes both the building's geometric form information and function type information. For each grid, the extended morphological feature vectors of the N building instances included in the grid are concatenated row by row to construct a building morphological feature matrix with a dimension of N×4. Each row of the matrix corresponds to a building instance, and the four columns correspond to the building height value, building base area value, building shape coefficient value, and building function type feature value, respectively. In subsequent steps, an N×4 building morphology feature matrix is ​​used to replace the N×3 matrix and input into the pre-trained carbon emission allocation network. For building instances with similar geometric shapes but different functional types, the carbon emission allocation network can assign different allocation weights based on the differences in functional type feature values, making the allocation results closer to the actual energy consumption distribution pattern.

[0036] In the above technical solution, the feature values ​​of building function type are extracted from the fused image data by using a large model of building form. These feature values ​​are then concatenated with the geometric feature vector to generate an extended morphological feature vector with a dimension of 4. This enables the carbon emission allocation network to simultaneously utilize the geometric form information and function type information of buildings for carbon emission allocation, distinguish the differences in energy consumption intensity per unit area of ​​buildings with different function types, and improve the spatial accuracy of carbon emission allocation in mixed-function areas.

[0037] Carbon emission allocation networks need to learn the mapping relationship from building morphological features to carbon emission allocation weights. Since true carbon emission data for each building is difficult to obtain in practice, the training of carbon emission allocation networks can only rely on the total grid carbon emissions as a weak supervision signal, leading to unreasonable weight allocation. In another technical solution, the carbon emission allocation network is trained using a weakly supervised learning framework, and the loss function consists of two parts: a total constraint term and a comparison constraint term. L = Lmse +λ×L contrast ; Among them, L mse The grid carbon emission total constraint term is calculated as the actual grid carbon emission total E. gt Carbon emissions are distributed among all buildings within the grid. i The mean square error of the sum, let the total carbon emissions of a certain grid be labeled as E. gt There are N building instances in the grid, and the weight assigned to the i-th building instance by the network output is w. i The base area is A i Then the carbon emission intensity c of this building example i = w i ×E gt / A i Since the sum of the allocation weights, after normalization, equals 1, the sum of the carbon emissions allocated to all buildings equals E. gt The theoretical minimum value of the total constraint term is zero, ensuring that the allocation result satisfies the basic physical requirement of total conservation. contrast To compare the constraints, the variance of carbon emission intensity of buildings of the same functional type within the same grid is calculated as the ratio of the variance of carbon emission intensity of buildings of different functional types. During training, for each grid, building instances are grouped according to their functional type labels. The functional type division is consistent with the building functional type feature values ​​output by the large building morphology model. The variance of carbon emission intensity of each building within each functional type group is calculated, and then the mean of the variances of each group is calculated as the within-group variance. Simultaneously, the variance between the means of carbon emission intensity of each group is calculated as the between-group variance. When the within-group variance is smaller and the between-group variance is larger, L... contrast The smaller the value, the more concentrated the carbon emission intensity of buildings of the same type and the more separated the carbon emission intensity of buildings of different types, and the more the allocation result matches the prior law. λ is the balance coefficient, which controls the importance of the comparison constraint term relative to the total constraint term. The value range is 0.1~0.5. The loss function guides the network to make the allocation result tend to a state in which the emission intensity of buildings of the same type is similar and the emission intensity of buildings of different types is significantly different, under the premise of satisfying the total constraint.

[0038] In the above technical solution, a contrastive constraint term is added to the loss function of the weakly supervised learning framework. The contrastive constraint term injects the general law of urban building energy consumption into the training process by minimizing the variance of carbon emission intensity of buildings with the same functional type and maximizing the variance of carbon emission intensity of buildings with different functional types, thereby suppressing the spurious correlation that may be generated by pure total constraint training.

[0039] The total carbon emissions values ​​in coarse-grained carbon emission grid data are typically regional total emissions, including non-building emission sources such as industry and transportation. This leads to a systematic overestimation of building-scale carbon emissions. In another technical approach, the total carbon emissions value for each grid is allocated according to the carbon emission distribution weights of each building instance within that grid. This process includes the following steps: Acquire land use classification data for the target area. The land use classification data includes building land, industrial land, transportation land, green space and water area. This data can be obtained from the results of the national land survey or remote sensing land cover classification products. Its spatial resolution should not be lower than the resolution of the coarse-grained carbon emission grid. The acquisition time should be consistent with the time reference of the carbon emission data. The surface carbon emission intensity of each land use type within the grid is positively correlated with its area ratio, meaning that the emissions per unit area of ​​industrial land are roughly comparable to those per unit area of ​​building land. Although the actual emission intensity varies among different land use types, in the absence of more refined emission source classification data, it is calculated based on area ratio. For each grid, calculate the proportion R of the built-up area to the total area of ​​the grid based on the land use classification data. build The value ranges from 0 to 1. This ratio is larger in urban grids with high building density and smaller in suburban or industrial grids. Multiply the total carbon emissions of this grid by R build This yields the total building carbon emissions for that grid. In subsequent steps, for each grid, its total building carbon emissions value is replaced with the original total carbon emissions value, and then allocated according to the carbon emissions allocation weight of each building instance within that grid.

[0040] In the above technical solution, by introducing land use classification data, the proportion of building land area in each grid to the total grid area is calculated, and the total carbon emissions of the grid are corrected to the total carbon emissions of buildings, thus avoiding interference from non-building emission sources such as industry and transportation.

[0041] The carbon emission allocation network outputs a set of defined allocation weights for each grid, and calculates the carbon emission estimate for each building instance based on this. Without verification of the true values ​​of emissions per building, the reliability of the allocation result cannot be known. In another technical solution, the carbon emission allocation network adopts the Monte Carlo dropout method during training and inference. A Dropout layer with a dropout rate of 0.1 to 0.2 is added after each fully connected layer. During the training phase, the Dropout layer randomly sets the output of some neurons to zero according to the set dropout rate. The output of the retained neurons is divided by 1 and subtracted from the dropout rate for scaling compensation. For the same grid of building morphology feature matrix, perform 20 to 50 forward propagations (without single forward propagation). In each forward propagation, the Dropout layer randomly discards different subsets of neurons. The network actually runs with slightly different parameter states, generating a set of carbon emission allocation weights that are slightly different from those of a single propagation. After multiple forward propagations, multiple sets of carbon emission allocation weights are obtained. For the t-th forward propagation, after obtaining a set of allocation weights, the carbon emissions allocated to each building instance are calculated by multiplying the weights by the total carbon emissions value of the grid. For the i-th building instance in the grid, let the allocation weight obtained in the t-th forward propagation be w. i (t), the total carbon emissions of the grid are E total The carbon emissions allocated to the building in this propagation are E. i (t)= E total × w i (t); After T propagations, the i-th building instance obtains T carbon emission samples. The mean μ of the carbon emissions allocated to each building instance corresponding to each weight group is calculated. i and standard deviation σ i : , , If the network's allocation weight for a building fluctuates significantly under different dropout modes, the standard deviation is large, indicating that the carbon emission estimation of that building has high uncertainty. If the allocation results of each propagation are relatively consistent, the standard deviation is small, indicating that the estimation results are relatively reliable. The mean and standard deviation of carbon emissions of each building instance are recorded simultaneously in the carbon emission spatial distribution data at the building scale to generate spatial distribution data with uncertainty of carbon emissions per building.

[0042] In the above technical solution, a Dropout layer is added after the fully connected layer of the carbon emission allocation network. The Monte Carlo dropout method is used to perform multiple forward propagations on the same grid to obtain multiple sets of allocation weights. The mean and standard deviation of carbon emissions for each building instance are statistically calculated. The mean is used as the carbon emission estimate and the standard deviation is used as the uncertainty measure. The generated building-scale carbon emission spatial distribution data is accompanied by building-by-building uncertainty information, providing a confidence reference for carbon emission management decisions.

[0043] A 5km × 5km built-up area in a northern city was selected as the test area. This area contains 812 buildings, including 512 residential buildings, 186 office buildings, and 114 commercial buildings. Optical remote sensing images with a spatial resolution of 0.8m and synthetic aperture radar images with a spatial resolution of 2m were acquired for this area. After preprocessing and fusion, fused image data was obtained. The fused image data was input into a large-scale building morphology model to extract the building height, building base area, and building shape coefficient of each building. At the same time, coarse-grained carbon emission grid data for this area was acquired. The grid side length is 500m, and a total of 100 grids are divided. Each grid records an annual total carbon emission value.

[0044] To obtain the benchmark value for carbon emissions per building, building energy consumption simulation software was used to simulate the energy consumption of each building in the test area. The thermal parameters of the building envelope were taken according to the building energy-saving design standards of the climate zone to which the test area belongs. The meteorological data used were typical meteorological data of the city. Geometric models were established for each building and an operation schedule was set. The annual operating energy consumption of each building was output. The annual operating energy consumption was multiplied by the local power grid carbon emission factor of 0.6105 kgCO2 / kWh to obtain the annual carbon emission benchmark value of each building, which served as the reference true value for accuracy assessment.

[0045] The carbon emission allocation network was pre-trained on three urban blocks that were not adjacent to the test area, with a total training data area of ​​approximately 120 km². 2 The test network includes a building morphology feature matrix of approximately 4,200 buildings. The grid data of the test area was not used in any form for training. After training, the building morphology feature matrix of the test area was input into the carbon emission allocation network. The Monte Carlo dropout method was used for 30 forward propagations to obtain the mean and standard deviation of the carbon emission allocation weights for each building instance. The total carbon emissions of the grid were allocated to each building according to the weight mean to generate a building-by-building carbon emission estimate. The average standard deviation of the carbon emission estimates for the 812 buildings in the test area was 5.2 kg CO2 / year.

[0046] A baseline method is established for comparison. The baseline method uses the proportion of each building's volume to the total volume of all buildings within its grid as the carbon emission allocation weight. Building volume is calculated as the product of the base area and the building height. The total carbon emissions of the grid are allocated to each building proportionally by volume. Both the proposed method and the baseline method are run separately under the same test area and grid data conditions, repeated five times with randomly divided grids for training and validation sets. The coefficient of determination R is then calculated. 2 The mean and standard deviation of the mean absolute percentage error (MAPE), and the performance comparison results between this method and the baseline method are shown in Table 1. The R-values ​​of this method and the baseline method grouped by building type are also shown in Table 1. 2 See Table 2 for comparison.

[0047] Table 1 Table 2 As shown in Table 1, the R of this method 2 The method improved by 43.1% compared to the baseline method, while MAPE decreased by 40.3%. Table 2 shows that this method improved R in residential, office, and commercial buildings. 2 All values ​​are higher than the baseline method, with the largest increase in office buildings. This is consistent with the high correlation between the energy consumption intensity per unit area of ​​office buildings and the shape coefficient. The above results show that the building shape coefficient and building function type characteristic value introduced in this method make the carbon emission allocation accuracy significantly better than the baseline method that only relies on volume weighting.

[0048] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention. Applications, modifications, and variations of the invention will be readily apparent to those skilled in the art.

[0049] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. Other modifications can be easily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details shown and described herein.

Claims

1. A carbon emission downscaling method based on multi-source remote sensing imagery and large-scale building morphology models, characterized in that, Includes the following steps: Acquire multi-source remote sensing images of the target area, including optical remote sensing images with a spatial resolution of 0.5 to 1 m and synthetic aperture radar images with a spatial resolution of 1 to 3 m. Perform radiometric calibration and atmospheric correction on the optical remote sensing images, perform multi-view processing on the synthetic aperture radar images, and register and fuse the processed images to obtain fused image data. The fused image data is input into a pre-trained large-scale building morphology model. The fused image data is segmented into building instances, and the building outline of each building instance and the morphological feature vector composed of building height, building base area and building shape coefficient are output. The coordinates of the center point of the bounding rectangle of each building outline are calculated as the geographical coordinates of each building instance. The large-scale building morphology model is a deep neural network trained with sample multi-source remote sensing images and corresponding labeled data of building height, building base area and building shape coefficient. Obtain coarse-grained carbon emission grid data for the target area. The grid cell side length of the coarse-grained carbon emission grid data is within 500~1000m, and each grid records a total carbon emission value. Based on the geographical coordinates of each building instance, all building instances whose center point coordinates fall into the same grid are grouped into the building instance set of that grid. For each grid, the morphological feature vectors of the N building instances included in the grid are concatenated row by row to construct a building morphological feature matrix of dimension N×3. The building morphological feature matrix is ​​input into a pre-trained carbon emission allocation network, which outputs the carbon emission allocation weight of each building instance in the grid, and satisfies that the sum of the carbon emission allocation weights of all building instances in the grid is 1. The carbon emission allocation network is a neural network. For each grid, its total carbon emissions are allocated according to the carbon emission allocation weights of each building instance within that grid, generating spatial distribution data of carbon emissions at the building scale.

2. The carbon emission downscaling method based on multi-source remote sensing imagery and large-scale building morphology model according to claim 1, characterized in that, The carbon emission allocation network is a neural network based on a deep set architecture. The deep set architecture includes a shared encoder network and a permutation-invariant aggregation function. The encoder network independently maps each row of the morphological feature vector in the N×3 building morphological feature matrix to a high-dimensional latent space. The aggregation function aggregates all the mapped latent vectors into a fixed-length global representation vector. Then, based on the global representation vector, the carbon emission allocation weight of each building instance is decoded and output to adapt to the change of the number of building instances N in the grid.

3. The carbon emission downscaling method based on multi-source remote sensing imagery and large-scale building morphology model according to claim 1, characterized in that, For building instances whose outlines cover more than two coarse-grained carbon emission grids, the area weight is the ratio of the area of ​​the intersection between the building instance and each grid to the area of ​​its building base. The carbon emission allocation weight of the building instance is then weighted and corrected using the area weight as a coefficient. After correction, the sum of the allocation weights of all building instances in each grid is 1, and then they participate in the allocation of the total carbon emission value in their respective grids.

4. The carbon emission downscaling method based on multi-source remote sensing imagery and large-scale building morphology model according to claim 1, characterized in that, The processed images are registered and fused to obtain fused image data, which includes: Building outlines are pre-extracted from the processed optical remote sensing images to generate binary optical building outline templates; Using the optical building outline template as a reference, the processed synthetic aperture radar image is matched and registered by outline feature matching. The RANSAC algorithm is used to remove mismatched point pairs so that the final registration error does not exceed 0.3 optical image pixels. The SwinFusion algorithm based on cross-modal attention mechanism is used to perform feature-level fusion of the registered images. In this algorithm, optical remote sensing image features are used as queries and synthetic aperture radar (SAR) image features are used as keys and values. Cross-modal feature interaction is achieved through cross-attention calculation. During the fusion process, an identity mapping is applied to the phase components of the SAR image features to generate fused image data.

5. The carbon emission downscaling method based on multi-source remote sensing imagery and large-scale building morphology model according to claim 1, characterized in that, The large model of the building form includes a 3D reconstruction sub-network. The 3D reconstruction sub-network takes fused image data as input. First, the monocular depth estimation module generates a disparity map and converts it into a normalized digital surface model. Then, the voxel reconstruction module outputs the voxel occupancy probability of the building instance. Based on the voxel occupancy probability, the ratio of the external surface area to the volume is calculated to obtain the building form coefficient value.

6. The carbon emission downscaling method based on multi-source remote sensing imagery and large-scale building morphology model according to claim 1, characterized in that, The large-scale building model uses SwinTransformer as the backbone network, with three parallel task heads set at the end: building instance segmentation head, building height regression head, and building shape coefficient regression head. The training loss function for the large-scale architectural model adopts a dynamically weighted multi-task loss function, and the calculation formula is as follows: L total = α(t)×L seg +β(t)×L height +γ(t)×L shape , Among them, L seg L is a weighted sum of the Dice loss and cross-entropy loss for the segmentation task. height For Huber loss in high regression tasks, L shape For the Huber loss in the body size coefficient regression task, α(t), β(t), and γ(t) are dynamic weights that are adaptively adjusted with the training epoch t. The adjustment rule for the dynamic weights is as follows: After every K training rounds, calculate the current stage L. seg L height L shape The respective rate of decline is calculated as the ratio of the reciprocal of the rate of decline of each loss in the current stage to the sum of the reciprocals of the three losses, and after being corrected by the smoothing coefficient τ, it is used as the dynamic weight of each loss in the next stage, where K takes the value of 5~10 and τ takes the value of 0.8~0.

95.

7. The carbon emission downscaling method based on multi-source remote sensing imagery and large-scale building morphology model according to claim 1, characterized in that, Before constructing the N×3 building morphology feature matrix, the following steps are also included: The architectural function type feature values ​​of each building instance are extracted from the fused image data using a large architectural morphology model. The building function type feature value is concatenated with the original form feature vector to generate an extended form feature vector with a dimension of 4. For each grid, the extended morphological feature vectors of the N building instances included in the grid are concatenated row by row to construct a building morphological feature matrix with a dimension of N×4. In subsequent steps, an N×4 building morphology feature matrix is ​​input into a pre-trained carbon emission allocation network.

8. The carbon emission downscaling method based on multi-source remote sensing imagery and large-scale building morphology model according to claim 1, characterized in that, The carbon emission allocation network is trained using a weakly supervised learning framework, and the loss function consists of two parts: a total constraint term and a comparison constraint term. L = L mse +λ×L contrast ; Among them, L mse For the total carbon emissions constraint term in the grid, L is calculated as the mean square error between the actual total carbon emissions of the grid and the sum of the carbon emissions allocated to all buildings within that grid. contrast To compare the constraint terms, the variance of carbon emission intensity of buildings of the same functional type within the same grid is calculated as the ratio of the variance of carbon emission intensity of buildings of different functional types. λ is the balance coefficient, with a value ranging from 0.1 to 0.

5.

9. The carbon emission downscaling method based on multi-source remote sensing imagery and large-scale building morphology model according to claim 1, characterized in that, For each grid, its total carbon emissions are allocated according to the carbon emission distribution weights of each building instance within that grid, preceded by the following steps: Obtain land use classification data for the target area, which includes building land, industrial land, transportation land, green space, and water area; For each grid, calculate the proportion R of the built-up area to the total area of ​​the grid based on the land use classification data. build ; Multiply the total carbon emissions of this grid by R build This yields the total building carbon emissions for that grid. In subsequent steps, for each grid, its total building carbon emissions value is allocated according to the carbon emission allocation weight of each building instance within that grid.

10. The carbon emission downscaling method based on multi-source remote sensing imagery and large-scale building morphology model according to claim 1, characterized in that, The carbon emission allocation network employs Monte Carlo dropout during training and inference, adding a Dropout layer with a dropout rate of 0.1 to 0.2 after each fully connected layer; By performing 20 to 50 forward propagations on the building morphology feature matrix of the same grid, multiple sets of carbon emission allocation weights are obtained. The mean and standard deviation of carbon emissions allocated to each building instance corresponding to each weight group are calculated to generate spatial distribution data with uncertainty in carbon emissions per building.

Images