A local climate zoning classification method based on real three-dimensional morphology of buildings
Patent Information
- Application Number
- CN202611175376.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-04
- Publication Date
- 2026-10-23
AI Technical Summary
遥感方法侧重地表光谱与纹理信息，对建筑垂直结构的表达能力有限；基于地理信息系统的方法通常将建筑高度、建筑密度和容积率等压缩为少量统计量，难以反映建筑个体轮廓、高低组合及空间布局差异
[0034]1．本发明突破了传统方法依赖二维影像或统计指标的局限，直接以建筑真实三维外轮廓点云作为分类依据，能够完整保留建筑个体形态、高度变化、边界转折及群体空间布局等三维结构信息，显著提升了局部气候分区分类的准确性。
Smart Images
Figure CN122967041A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of geographic information processing technology, and more specifically, to a method for classifying local climate zones based on the actual three-dimensional form of buildings. Background Technology
[0002] Local climate zoning is a standardized classification system commonly used in urban thermal environment research. It can uniformly divide urban space based on indicators such as building height, building density, and land cover, thus providing a basis for comparative analysis of thermal environments between different cities. Urban blocks constitute the basic unit of local climate zoning research. The three-dimensional morphology of buildings within a block has a direct impact on local ventilation, radiation, and heat exchange processes, and is also an important basis for distinguishing various local climate zones.
[0003] Current methods for obtaining local climate zones mostly rely on remote sensing imagery or street statistical indicators. Remote sensing methods focus on surface spectral and textural information, but have limited ability to express the vertical structure of buildings; Geographic Information System (GIS)-based methods usually compress building height, building density, and floor area ratio into a small number of statistical quantities, which are difficult to reflect the individual building outlines, height combinations, and spatial layout differences.
[0004] For types of high-rise buildings, such as compact high-rises and open high-rises, which have similar statistical indicators but different spatial organization methods, relying solely on two-dimensional images or statistical values can easily lead to category confusion. 3D point clouds of building outlines can directly record structural information such as the building's bottom, top, boundaries, edges, and corners; however, the point distribution after converting the original vectors to points is usually sparse, making it difficult to stably support deep networks in extracting local geometric features.
[0005] Therefore, it is necessary to construct a method that can maintain the true three-dimensional outline of buildings, enhance the expression of boundaries and structural transitions, and use point cloud deep learning to achieve local climate zoning classification at the street scale. Summary of the Invention
[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for classifying local climate zones based on the actual three-dimensional shape of buildings.
[0007] Firstly, a method for classifying local climate zones based on the actual three-dimensional form of buildings is provided, including:
[0008] S1. Collect street vector data and building vector data of the target city, and correct the street boundaries, building outlines and height attributes; label the street samples according to the local climate zoning rules to determine the classification target category;
[0009] S2. Stretch the corrected building vector along the vertical direction according to the height attribute to construct a three-dimensional building block model; extract the three-dimensional coordinates of the building's bottom contour vertices, top contour vertices, and vertical boundary points, and organize them into a building outer contour point cloud according to the street number.
[0010] S3. Using the three-dimensional outline of the building as a spatial constraint, the initial point cloud is upsampled so that the newly sampled points are distributed at the building boundary, edge, and corner positions, thus obtaining an optimized point cloud that maintains the true outer outline structure of the building.
[0011] S4. Input the optimized point cloud into a dynamic graph convolutional neural network, extract multi-scale geometric features by dynamically constructing a local neighborhood graph, form street-level features through feature splicing and pooling, and output the local climate zone category using a classifier.
[0012] As a preferred embodiment, S1 specifically includes:
[0013] S101. Combine high-resolution imagery and road network data to perform unit-by-unit correction of block boundaries, and check and correct building outlines and height attributes building by building;
[0014] S102. Organize labeling personnel to determine the street type based on building height, building density and spatial layout, and correct inconsistent samples through cross-verification;
[0015] S103. Refine the blocks with significant differences in internal building forms and divide them into training and test sets according to a set ratio.
[0016] As a preferred option, S2 specifically includes:
[0017] S201. Establish a bottom contour point set with a height of 0 and a top contour point set with a height equal to the building height for each building, and establish an upper and lower correspondence relationship according to the order of the same contour vertices.
[0018] S202. Merge the bottom contour points, top contour points, and the vertical boundary points connecting the two, and organize them into a point cloud tensor according to the street number.
[0019] Preferably, S3 includes:
[0020] S301. Using the bottom boundary, top boundary, and vertical boundary of the building as the allowed sampling domain, generate supplementary points between adjacent structural points so that the newly added points are strictly located on the three-dimensional outer contour of the building.
[0021] S302. Set a uniform number of points for each block, perform spatial constraint upsampling for blocks with insufficient points, and perform downsampling to maintain the boundary distribution for blocks with more than the set number of points.
[0022] Preferably, in S4, the dynamic graph convolutional neural network includes a multi-layer dynamic graph convolutional module, where each layer dynamically establishes a local graph based on k nearest neighbors, and the output dimension increases sequentially; the outputs of each layer are concatenated along the channel direction to form point-level features, and block-level features are obtained through global pooling and input into a fully connected classification head.
[0023] Preferably, the edge features of the dynamic graph convolution module are composed of the center point features and the relative features of the neighboring points, and the neighborhood edge features are subjected to dimension-wise max pooling to update the center point features.
[0024] Secondly, a local climate zoning classification system based on the actual three-dimensional form of a building is provided for performing any of the methods described in the first aspect, including:
[0025] The data acquisition module is used to collect vector data of blocks and buildings in the target city, correct the block boundaries, building outlines and height attributes; and label the block samples according to the local climate zoning rules to determine the classification target category.
[0026] The building module is used to stretch the corrected building vectors along the vertical direction according to the height attribute to build a three-dimensional building block model; extract the three-dimensional coordinates of the bottom contour vertices, top contour vertices and vertical boundary points of the building, and organize them into a building outer contour point cloud according to the block number;
[0027] The sampling module is used to upsample the initial point cloud using the three-dimensional outline of the building as a spatial constraint, so that the newly added sampling points are distributed at the building boundary, edges and corners, and an optimized point cloud that maintains the true outer outline structure of the building is obtained.
[0028] The output module is used to input the optimized point cloud into a dynamic graph convolutional neural network, extract multi-scale geometric features by dynamically constructing a local neighborhood graph, form street-level features through feature concatenation and pooling, and output the local climate zone category using a classifier.
[0029] Thirdly, a computer storage medium is provided, wherein a computer program is stored therein; when the computer program is run on a computer, the computer causes the computer to perform any of the methods described in the first aspect.
[0030] Fourthly, an electronic device is provided, comprising:
[0031] Memory, used to store computer programs;
[0032] A processor for executing the computer program to implement the method as described in any of the first aspects.
[0033] The beneficial effects of this invention are:
[0034] 1. This invention breaks through the limitations of traditional methods that rely on two-dimensional images or statistical indicators. It directly uses the real three-dimensional outline point cloud of buildings as the classification basis, which can completely preserve the three-dimensional structural information such as the shape of individual buildings, height changes, boundary transitions and group spatial layout, and significantly improve the accuracy of local climate zone classification.
[0035] 2. This invention proposes a point cloud optimization strategy based on the three-dimensional contour of a building as a spatial constraint. This strategy ensures that newly added sampling points are strictly distributed at the bottom boundary, top boundary, vertical boundary, edges, and corners of the building. This guarantees high-density representation of the point cloud in areas with drastic geometric changes, providing high-quality data input with intact structure for subsequent deep learning models.
[0036] 3. This invention uses a dynamic graph convolutional neural network to classify building point clouds. Through the dynamic graph construction and edge feature aggregation mechanism of the network layers, multi-scale geometric features from local structure to global topology are extracted layer by layer, realizing deep modeling of the spatial relationship of building groups, and further improving the reliability and generalization ability of local climate zone classification. Attached Figure Description
[0037] Figure 1 This is a flowchart of data acquisition, correction, and sample construction according to an embodiment of the present invention;
[0038] Figure 2 This is a flowchart of the construction of a real 3D point cloud of a building and the upsampling of spatial constraints according to an embodiment of the present invention;
[0039] Figure 3 This is a flowchart of the local climate zoning classification based on dynamic graph convolutional neural network according to an embodiment of the present invention. Detailed Implementation
[0040] The present invention will be further described below with reference to embodiments. The description of the embodiments below is only for the purpose of helping to understand the present invention. It should be noted that those skilled in the art can make several modifications to the present invention without departing from the principle of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.
[0041] Example 1:
[0042] To address the problems of existing technologies, Embodiment 1 of this application provides a method for classifying local climate zones based on the actual three-dimensional form of buildings, including:
[0043] S1. Collect street vector data and building vector data of the target city, and correct the street boundaries, building outlines and height attributes; label the street samples according to the local climate zoning rules to determine the classification target category.
[0044] S1 specifically includes:
[0045] S101. Combine high-resolution imagery and road network data to perform unit-by-unit correction of street boundaries, and check and correct building outlines and height attributes building by building.
[0046] Specifically, global multi-scale street block division data and building vector data of the target city were collected. The street block data used the L5 level division results of the MSDCW dataset, with the finest-grained street blocks formed by the natural segmentation of the urban road network as the classification unit. Then, the street block boundaries were verified unit by unit by combining high-resolution imagery and OSM road network, correcting road network offsets, incorrect street block merging, and deviations in built-up area boundaries; the building outlines were checked building by building with high-resolution imagery as a reference, missing elements were supplemented, and building height attributes were corrected.
[0047] S102. The organization of labeling personnel completes the determination of street type based on building height, building density and spatial layout, and corrects inconsistent samples through cross-verification.
[0048] Specifically, based on the built environment characteristics of the study area, five types of buildings—compact high-rise, compact mid-rise, open high-rise, open mid-rise, and large low-rise—were selected for local climate zones. The annotation personnel first studied the standard definitions and typical samples, then made judgments based on building height, density, and spatial layout, and finally corrected inconsistencies through cross-verification.
[0049] S103. Refine the blocks with significant differences in internal building forms and divide them into training and test sets according to a set ratio.
[0050] Specifically, for blocks with large areas and inconsistent building group shapes, Voronoi diagrams are constructed by extracting building outline vertices and then refined based on the building group shape boundaries. Stratified random sampling is used to divide the training and test sets in an 8:2 ratio.
[0051] S2. Stretch the corrected building vector along the vertical direction according to the height attribute to construct a three-dimensional building block model; extract the three-dimensional coordinates of the bottom contour vertices, top contour vertices, and vertical boundary points of the building, and organize them into a point cloud of the building's outer contour according to the street number.
[0052] S2 specifically includes:
[0053] S201. For each building, establish a bottom contour point set with a height of 0 and a top contour point set with a height equal to the building height, and establish an upper and lower correspondence relationship according to the order of the same contour vertices.
[0054] Specifically, the building vector data is copied, and the height attribute of the copied data is set to 0. The original building outline and the building outline with a height of 0 are then extracted. The points with a height of 0 constitute the bottom outline of the building, and the original height points constitute the top outline of the building.
[0055] S202. Merge the bottom contour points, top contour points, and the vertical boundary points connecting the two, and organize them into a point cloud tensor according to the street number.
[0056] Specifically, the top and bottom contour points are matched according to the building number, and vertical boundary points are established between the corresponding vertices. The bottom contour, top contour, and vertical boundary are merged to form a three-dimensional point set that can represent the structure of the building's outer surface. Then, the point clouds of all buildings within the block are compiled according to the block number, the X, Y, and Z coordinates of each point are recorded, and the block point set is organized into an N×3 input tensor.
[0057] S3. Using the three-dimensional outline of the building as a spatial constraint, the initial point cloud is upsampled so that the newly added sampling points are distributed at the building boundary, edge and corner positions, thus obtaining an optimized point cloud that maintains the true outer outline structure of the building.
[0058] S4. Input the optimized point cloud into a dynamic graph convolutional neural network, extract multi-scale geometric features by dynamically constructing a local neighborhood graph, form street-level features through feature splicing and pooling, and output the local climate zone category using a classifier.
[0059] Example 2:
[0060] Based on Example 1, Example 2 of the present invention provides a more specific method for classifying local climate zones based on the actual three-dimensional form of buildings, including:
[0061] S1. Collect street vector data and building vector data of the target city, and correct the street boundaries, building outlines and height attributes; label the street samples according to the local climate zoning rules to determine the classification target category.
[0062] S2. Stretch the corrected building vector along the vertical direction according to the height attribute to construct a three-dimensional building block model; extract the three-dimensional coordinates of the bottom contour vertices, top contour vertices, and vertical boundary points of the building, and organize them into a point cloud of the building's outer contour according to the street number.
[0063] S3. Using the three-dimensional outline of the building as a spatial constraint, the initial point cloud is upsampled so that the newly added sampling points are distributed at the building boundary, edge and corner positions, thus obtaining an optimized point cloud that maintains the true outer outline structure of the building.
[0064] S3 includes:
[0065] S301. Using the bottom boundary, top boundary, and vertical boundary of the building as the allowed sampling domain, supplementary points are generated between adjacent structural points to ensure that the newly added points are strictly located on the three-dimensional outer contour of the building, avoiding points that deviate from the real building contour generated by ordinary random interpolation.
[0066] S302. Using a single building as the processing unit, spatial constraint upsampling is performed on the 3D structural point cloud of the building, and the points are aggregated according to the street block number to form a street block building point cloud. This spatial constraint upsampling further increases the local sampling density by retaining the existing bottom contour points, top contour points, and vertical boundary points in the building point cloud before upsampling—these so-called original structural points—to improve the completeness of the representation of building contours, corners, vertical edges, and areas of height variation. However, since the initial building point cloud is relatively sparse, downsampling might result in the loss of key boundary points and structural turning points, weakening the ability to express the true 3D form of the building. Therefore, to retain the actual number of points in each street block point cloud to the greatest extent possible, fixed point count processing and downsampling are not performed.
[0067] Specifically, spatial constraint upsampling uses a single building as the processing unit. First, a 5-nearest neighbor graph without self-loops is constructed within the point set of that building to avoid generating supplementary points between different buildings. For each original structural point... Randomly select 1 neighboring point from its 5 neighboring points. And a supplementary point is generated using linear interpolation. The formula is as follows:
[0068] (1)
[0069] Interpolation coefficients exist The points are randomly selected uniformly within the range; the upsampling factor is set to 2, so each original point with an effective neighborhood generates one supplementary point, increasing the number of points for a single building to approximately twice the original. Building corners are directly represented by the original vertices of the bottom and top contours, edges are represented by the vertical boundary points between the corresponding upper and lower contour vertices, and boundary changes are represented by the aforementioned structural points and their local 5 nearest neighbor relationships; this embodiment does not set a curvature threshold or corner detector for secondary recognition. After upsampling, the point cloud can more fully express the building contours, height changes, and the relationship between the building group's volume.
[0070] S4. Input the optimized point cloud into a dynamic graph convolutional neural network, extract multi-scale geometric features by dynamically constructing a local neighborhood graph, form street-level features through feature splicing and pooling, and output the local climate zone category using a classifier.
[0071] In S4, the dynamic graph convolutional neural network includes a multi-layer dynamic graph convolutional module. Each layer dynamically establishes a local graph based on k nearest neighbors, and the output dimension increases sequentially. The outputs of each layer are concatenated along the channel direction to form point-level features, and block-level features are obtained through global pooling and input into a fully connected classification head.
[0072] The edge features of the dynamic graph convolution module are composed of the center point features and the relative features of the neighboring points, and the neighborhood edge features are subjected to dimension-wise max pooling to update the center point features.
[0073] Specifically, the edge features of dynamic graph convolution are calculated as follows:
[0074] (2)
[0075] (3)
[0076] in, Indicates from the center point With neighboring points The edge features formed For parameters nonlinear mapping function, Features of the center point Features of neighboring points This represents the characteristic differences between neighboring points and the center point.
[0077] (4)
[0078] in, This represents the updated point features. For point The neighborhood set, This indicates that the neighborhood edge features are subjected to dimension-wise max pooling. Equation (3) is the neighborhood aggregation result of Equation (2): first, for the center point... With each of its neighboring points Calculate the edge features according to the first formula. ; then in Within the scope of all Dimensional max pooling is performed to obtain the updated center point features. Therefore, equation (2) defines the features of a single neighborhood edge, and equation (3) aggregates the features of multiple edges corresponding to the same center point into new point features.
[0079] Furthermore, S4 includes:
[0080] S401: The model input is an N×3 block building point cloud, and the output is a five-dimensional classification vector corresponding to five local climate zones. Each layer reconstructs a k-nearest neighbor graph based on the current features, with the nearest neighbor number k set to 20.
[0081] The S402 four-layer dynamic graph convolution module outputs dimensions of 64, 64, 128, and 256 respectively, with batch normalization and ReLU activation applied to each layer. The first and second layers focus on extracting local contours and height differences, while the third and fourth layers aggregate the spatial relationships and volumetric features of building groups.
[0082] S403. The outputs of the four layers are concatenated into 512-dimensional point-level features, and then global max pooling is used to obtain street-level features. These features are then input into a two-layer fully connected classification head. The random inactivation rate of the classification head is 0.5, the loss function is cross-entropy, the optimizer is Adam, the initial learning rate is 0.001, and the weight decay is 0.0005.
[0083] The embodiments of this invention also provide basic classification experimental results and comparative analysis, as shown in Table 1. Using a test set of 388 samples to validate the model, the overall accuracy of the basic classification method of this invention reached 85.25%, with a Kappa coefficient of 0.82. This method improves the overall accuracy by 21.09% compared to the optimal method based on overall street statistical features (GraphSAGE), by 17.47% compared to the optimal method based on three-dimensional building statistical features (GraphSAGE), and by 2.34% compared to the classic point cloud classification method PointNet++.
[0084] Table 1 Comparison of accuracy between basic classification methods and existing technologies.
[0085]
[0086] To verify the effectiveness of sample splitting and spatial constraint upsampling, four ablation experiments were conducted, and the results are shown in Table 2. When using only the original data, the accuracy of DGCNN was only 50.68%; when both sample splitting and spatial constraint upsampling were enabled, the accuracy improved to 85.25%, demonstrating that both modules significantly contribute to the classification performance.
[0087] Table 2 Ablation Experiment Results of Data Processing Module
[0088]
[0089] It should be noted that the parts in this embodiment that are the same as or similar to those in Embodiment 1 can be referred to each other, and will not be repeated in this application.
[0090] Example 3:
[0091] Based on Example 2, Example 3 of the present invention provides a local climate zoning classification system based on the actual three-dimensional form of a building, including:
[0092] The data acquisition module is used to collect vector data of blocks and buildings in the target city, correct the block boundaries, building outlines and height attributes; and label the block samples according to the local climate zoning rules to determine the classification target category.
[0093] The building module is used to stretch the corrected building vectors along the vertical direction according to the height attribute to build a three-dimensional building block model; extract the three-dimensional coordinates of the bottom contour vertices, top contour vertices and vertical boundary points of the building, and organize them into a building outer contour point cloud according to the block number;
[0094] The sampling module is used to upsample the initial point cloud using the three-dimensional outline of the building as a spatial constraint, so that the newly added sampling points are distributed at the building boundary, edges and corners, and an optimized point cloud that maintains the true outer outline structure of the building is obtained.
[0095] The output module is used to input the optimized point cloud into a dynamic graph convolutional neural network, extract multi-scale geometric features by dynamically constructing a local neighborhood graph, form street-level features through feature concatenation and pooling, and output the local climate zone category using a classifier.
[0096] It should be noted that the system provided in this embodiment is the system corresponding to the method provided in embodiment 2. Therefore, the parts that are the same as or similar to those in embodiment 2 in this embodiment can be referred to each other, and will not be described again in this application.
Claims
1. A method for classifying local climate zones based on the actual three-dimensional form of buildings, characterized in that, include: S1. Collect street vector data and building vector data of the target city, and correct the street boundaries, building outlines and height attributes; label the street samples according to the local climate zoning rules to determine the classification target category; S2. Stretch the corrected building vector along the vertical direction according to the height attribute to construct a three-dimensional building block model; extract the three-dimensional coordinates of the building's bottom contour vertices, top contour vertices, and vertical boundary points, and organize them into a building outer contour point cloud according to the street number. S3. Using the three-dimensional outline of the building as a spatial constraint, the initial point cloud is upsampled so that the newly sampled points are distributed at the building boundary, edge, and corner positions, thus obtaining an optimized point cloud that maintains the true outer outline structure of the building. S4. Input the optimized point cloud into a dynamic graph convolutional neural network, extract multi-scale geometric features by dynamically constructing a local neighborhood graph, form street-level features through feature splicing and pooling, and output the local climate zone category using a classifier.
2. The method for classifying local climate zones based on the true three-dimensional morphology of buildings according to claim 1, characterized in that, S1 specifically includes: S101. Combine high-resolution imagery and road network data to perform unit-by-unit correction of block boundaries, and check and correct building outlines and height attributes building by building; S102. Organize labeling personnel to determine the street type based on building height, building density and spatial layout, and correct inconsistent samples through cross-verification; S103. Refine the blocks with significant differences in internal building forms and divide them into training and test sets according to a set ratio.
3. The method for classifying local climate zones based on the true three-dimensional morphology of buildings according to claim 2, characterized in that, S2 specifically includes: S201. Establish a bottom contour point set with a height of 0 and a top contour point set with a height equal to the building height for each building, and establish an upper and lower correspondence relationship according to the order of the same contour vertices. S202. Merge the bottom contour points, top contour points, and the vertical boundary points connecting the two, and organize them into a point cloud tensor according to the street number.
4. The method for classifying local climate zones based on the true three-dimensional form of buildings according to claim 3, characterized in that, S3 include: S301. Using the bottom boundary, top boundary, and vertical boundary of the building as the allowed sampling domain, generate supplementary points between adjacent structural points so that the newly added points are strictly located on the three-dimensional outer contour of the building. S302. Set a uniform number of points for each block, perform spatial constraint upsampling for blocks with insufficient points, and perform downsampling to maintain the boundary distribution for blocks with more than the set number of points.
5. The method for classifying local climate zones based on the true three-dimensional morphology of buildings according to claim 4, characterized in that, In S4, the dynamic graph convolutional neural network includes a multi-layer dynamic graph convolutional module. Each layer dynamically establishes a local graph based on k nearest neighbors, and the output dimension increases sequentially. The outputs of each layer are concatenated along the channel direction to form point-level features, and block-level features are obtained through global pooling and input into a fully connected classification head.
6. The method for classifying local climate zones based on the true three-dimensional morphology of buildings according to claim 5, characterized in that, The edge features of the dynamic graph convolution module are composed of the center point features and the relative features of the neighboring points, and the neighborhood edge features are subjected to dimension-wise max pooling to update the center point features.
7. A local climate zoning classification system based on the actual three-dimensional form of buildings, characterized in that, For performing the method as described in any one of claims 1 to 6, comprising: The data acquisition module is used to collect vector data of blocks and buildings in the target city, correct the block boundaries, building outlines and height attributes; and label the block samples according to the local climate zoning rules to determine the classification target category. The building module is used to stretch the corrected building vectors along the vertical direction according to the height attribute to build a three-dimensional building block model; extract the three-dimensional coordinates of the bottom contour vertices, top contour vertices and vertical boundary points of the building, and organize them into a building outer contour point cloud according to the block number; The sampling module is used to upsample the initial point cloud using the three-dimensional outline of the building as a spatial constraint, so that the newly added sampling points are distributed at the building boundary, edges and corners, and an optimized point cloud that maintains the true outer outline structure of the building is obtained. The output module is used to input the optimized point cloud into a dynamic graph convolutional neural network, extract multi-scale geometric features by dynamically constructing a local neighborhood graph, form street-level features through feature concatenation and pooling, and output the local climate zone category using a classifier.
8. A computer storage medium, characterized in that, The computer storage medium stores a computer program; when the computer program is run on the computer, it causes the computer to perform the method described in any one of claims 1 to 6.
9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the method as described in any one of claims 1 to 6.