Building knowledge graph construction and aggregation method for multi-scale expression

Abstract

A building knowledge graph construction and aggregation method for multi-scale expression comprises the following steps: step 1: obtaining building vector map data, performing data preprocessing, and forming a standardized building basic data set; step 2: based on the building basic data set, spatial and semantic relationship modeling is performed to construct a building level basic graph structure; step 3: rule pruning and graph theory clustering are performed on the constructed basic graph structure to form a building group unit to construct a single-layer building knowledge graph; step 4: land use category reasoning is performed by fusing spatial constraint information and multi-source geographic features to establish the correlation between different levels and construct a multi-scale urban building knowledge graph. The present application starts from the building basic unit, constructs a micro-meso-macro multi-level knowledge graph structure, enhances the expression ability of building correlation, and further improves the explainability, expansibility and applicability of urban land use mapping results.

Description

Technical Field

[0001] This invention relates to computer technology and cartographic technology, and more particularly to map making and knowledge graph reasoning technology, specifically to a method for constructing and aggregating a building knowledge graph for multi-scale expression. Background Technology

[0002] Urban land use, as a crucial spatial carrier reflecting the interaction between human activities and the natural environment, has become a fundamental data support for fields such as land spatial planning, urban governance, ecological environment monitoring, and public service optimization. Buildings, as basic units constituting urban space, carry multi-dimensional information including functional attributes, geometric forms, and property rights. Their organizational patterns and aggregation logic at different scales directly affect the spatial accuracy and semantic consistency of urban multi-scale representation. In the process of building mapping, urban land use spatial representation exhibits significant scale-dependent characteristics. Specifically, at the macro scale, mapping focuses on describing the overall distribution and spatial structure of buildings; at the meso scale, the core of mapping shifts to functional zoning identification and the internal organizational relationships of blocks; and at the micro-scale, the mapping content needs to deeply depict the geometric outlines, distribution density, ownership boundaries, and inter-unit relationships of buildings. Therefore, multi-scale building mapping is not a simple process of scaling up or simplifying geometry, but rather requires a targeted and comprehensive expression of the geometric structural characteristics, spatial topological relationships, and semantic attribute information of buildings at different scales.

[0003] In recent years, research on urban building mapping has gradually evolved from single-data-source, single-granularity methods to methods that integrate multi-source data and support multi-granularity representation. Early multi-scale building mapping methods, in the process of cross-scale representation, focused primarily on recognizing spatial distribution patterns of building clusters. The technical approach has generally evolved from traditional rule-based methods to machine learning and deep learning. These methods mainly employ template matching, nearest-neighbor graph construction, minimum spanning tree pruning, stroke constraints, and relative nearest-neighbor graphs to identify linear and grid-like patterns of buildings. These methods have clear rules and strong interpretability, but their adaptability to complex shapes, local heterogeneity, and large-scale scenes is limited. Wei Zhiwei et al. further proposed a method combining Gestalt principles and graph convex decomposition to improve the fragmentation and omission problems caused by treating only the building as a whole as a cognitive unit; subsequently, knowledge graphs and rule-based reasoning were introduced to improve pattern recognition efficiency. At the same time, machine learning and deep learning methods such as random forests and graph convolutional neural networks have also been used for building cluster pattern recognition, showing certain advantages in representing complex features. However, research by Tang Zengyang et al. and Xie Mengyuan et al. indicates that existing methods are still insufficient for the unified organization and collaborative reasoning of multi-level information such as spatial relationships, semantic attributes, and group patterns, and cannot yet meet the needs of multi-level information collaborative expression and interpretable reasoning in building aggregation mapping. At the building data organization level, knowledge graph technology has attracted attention due to its powerful semantic expression and relational reasoning capabilities. Some studies have attempted to construct geographic knowledge graphs or urban knowledge graphs, structurally expressing geographic entities such as buildings, roads, plots, and points of interest, as well as their spatial and semantic relationships, to support urban function identification, spatial querying, and intelligent analysis. At the building aggregation level, existing methods are mostly based on geometric proximity, morphological similarity, or road network constraints, using techniques such as clustering, region growing, or graph segmentation to merge fine-grained buildings into higher-level spatial units for street division, urban structure analysis, or map generalization. Despite the progress made in the above research, existing technologies still face two key problems that urgently need to be addressed.

[0004] On the one hand, most architectural drawing representations focus on the organization of building entity relationships, structural pattern recognition, and geometric feature transformation at a single scale, lacking explicit modeling of the aggregation relationships of buildings at different scales, making it difficult to directly support the flexible generation and dynamic updating of multi-scale building representations.

[0005] On the other hand, existing building merging methods mostly rely on simple geometric and topological features, rarely incorporating high-level semantic information such as building function as constraints into the merging decision-making process. This results in insufficient consistency in the functional semantics of the merging results, making it difficult to effectively serve cross-scale urban analysis and planning applications. In fact, buildings within the same land use type often have functional homogeneity and tend to form aggregated units in space; while the boundaries between different land use types set semantic constraints for building merging.

[0006] Therefore, how to integrate the spatial relationships, geometric features, and high-level semantic information such as land use of buildings within the framework of geographic knowledge graphs, and how to construct multi-scale merging rules that combine graphics and relational constraints on this basis, has become a key issue in improving the multi-scale representation and mapping of buildings. Summary of the Invention

[0007] The purpose of this invention is to address the following technical problems existing in the multi-scale organization and aggregation of buildings: aggregation relies solely on road network division or geometric proximity, ignoring the semantic relationships between buildings based on land use and functional attributes, resulting in insufficient semantic consistency and spatial relevance of the merged results; and the semantic utilization of multi-source information such as land use and points of interest is insufficient during the aggregation decision-making process. Therefore, this invention proposes a building knowledge graph construction and aggregation method oriented towards multi-scale expression. It aims to construct a multi-level knowledge graph structure (micro-meta-macro) starting from basic building units through the structured expression and rule-based reasoning mechanism of the knowledge graph, thereby enhancing the expressive power of building relationships and improving the interpretability, scalability, and applicability of urban land use mapping results.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for constructing and aggregating building knowledge graphs for multi-scale representation, comprising the following steps: Step 1: Obtain building vector map data, perform data preprocessing, and form a standardized building basic dataset; Step 2: Based on the building dataset, perform spatial and semantic relationship modeling to construct the basic graph structure of building hierarchy; Step 3: Perform rule-based pruning and graph theory clustering on the constructed basic graph structure to form building group units, thereby constructing a single-story building knowledge graph; Step 4: Integrate spatial constraint information with multi-source geographical features to infer land use categories, establish connections between different levels, and construct a multi-scale urban building knowledge graph.

[0009] Step 1 specifically includes the following sub-steps: Step 1-1) Collect building vector map data. The building vector map data is selected from regional scenes with high building density, complete road network structure and at least one urban functional space intertwined with commerce, residence, public services and industry. Steps 1-2) Perform data cleaning, geometric correction, attribute completion, semantic improvement, and category system alignment on the collected building vector map data to ensure that the buildings meet the requirements for subsequent processing in terms of spatial location, geometric integrity, and attribute information consistency, and finally form a standardized building basic dataset.

[0010] Step 2 specifically includes the following sub-steps: Step 2-1) Abstract each building entity into a graph node and bind it with the building name, function category, center point coordinates and geometric description attributes extracted in the preprocessing stage to construct the entity representation structure of the building hierarchy; Step 2-2) Construct a Delaunay triangulation based on the building center point, capture the spatial proximity and topological association between buildings, and define the building pairs with direct connection relationships in the triangulation as candidate spatial adjacency relationships; Steps 2-3) Combining road network information, the candidate spatial adjacency relationships are further subdivided into general adjacency relationships and cross-road adjacency relationships; Steps 2-4) use fuzzy string matching to calculate building name similarity, and combine this with the reclassified building function category system to comprehensively determine the functional consistency and name attribute correlation between buildings, constructing semantic similarity relationships between buildings, forming a basic hierarchical graph structure of buildings, formally represented as: ; in, Represents the set of entity nodes in a building layer. This represents the set of edges representing relationships between entities.

[0011] In step 3, performing rule-based pruning specifically includes the following sub-steps: Step 3-1) Perform the first step of rule pruning for building pairs that simultaneously have cross-road adjacency and semantic similarity relationships: when a building i With buildings j semantic similarity between When the value is below a preset first threshold of 0.9, its cross-road adjacency relationship is removed. The retained state after pruning is expressed as follows: ; in, i and j These represent the node numbers of any two buildings; Represents buildings iWith buildings j Whether there is a cross-road relationship between them, a value of 1 indicates that it exists, and a value of 0 indicates that it does not exist; Represents buildings i With buildings j Whether there is a semantic similarity relationship between them, a value of 1 indicates that there is, and a value of 0 indicates that there is no; Represents buildings i With buildings j The semantic similarity between the function and the name; and These represent the preserved states of cross-road relationships and semantic similarity relationships after the first pruning step, respectively. Step 3-2) Perform the second step of rule pruning for building pairs that simultaneously have both general adjacency and semantic similarity relationships: When buildings have both adjacency and semantic similarity relationships, prioritize retaining the adjacency relationship and delete the semantic similarity relationship to reduce redundancy. The retained state is expressed as: ; in, Represents buildings i With buildings j Whether there exists a general adjacency relationship between them that is directly adjacent in space, a value of 1 indicates that they exist, and a value of 0 indicates that they do not exist. and These represent the preserved states of general adjacency relations and semantic similarity relations after the second pruning step; " indicates a pruning logic mapping operation; Step 3-3) Introduce building shape similarity as a constraint for the third step of rule pruning. The calculation formula is as follows: ; in, A comprehensive index representing the morphological similarity between buildings. , and These represent the similarity of the corresponding areas, aspect ratios, and orientation angles of the two buildings, respectively.

[0012] In step 3, forming the building group unit specifically includes the following sub-steps: Step s3-1) Using the connected component identification method in graph theory, the building nodes corresponding to the relation edges retained after rule pruning are aggregated to form building group units with strong spatial-semantic consistency; Step s3-2) Based on the spatial location matching relationship between the building group clustering results and the AOI formed by road expansion, establish the inclusion relationship, and according to the real topological associations retained after rule pruning, merge multiple AOIs in the same connected component into the same block functional area unit, which is defined as the Landuse unit.

[0013] In step 4, land use category reasoning is performed on the block functional area unit Landuse, which specifically includes the following sub-steps: Step 4-1) Statistically analyze the proportion characteristics of each building at the first-level within the functional unit of the street block. Characteristics of the total number of buildings and category richness features , constituting a plot of land i Node feature vectors ; Initial pseudo-labels are generated using a majority voting method. and its confidence level ,in Indicates land parcel Belongs to category The number of buildings; in, K This indicates the total number of pre-defined primary building categories; k and c Index representing the category; Indicates land parcel i The number of buildings belonging to the kth first-level building category; This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. Step 4-2) Establish spatial adjacency constraints based on whether the plots share buildings, and construct a functional area association graph of the block. The spatial connection edges of its topological connection matrix are: ; Step 4-3) Input the node feature vectors and association graph into the network, and calculate the unnormalized attention coefficients. Normalized attention weights The hidden representation is obtained after feature aggregation. Output the basic prediction probability distribution ; in, The attention parameter weight vector is represented by W; the trainable linear transformation feature map matrix is ​​represented by ||; and the vector concatenation operator is represented by ||. Represents a node i The set of neighborhood nodes in the association graph Represents a node i Neighborhood nodes in the association graph j eigenvectors, Represents a nonlinear activation function; This leads to the basic land use inference results. and prediction confidence ;in, c Indicates category index, C This represents the set of all candidate land use categories. Represents the basic prediction probability distribution The central plot belongs to the candidate land use category c The probability value; Step 4-4) Calculate the category proportion vector at the first level of POI. ,in Represents land parcels Inner The proportion of POIs is defined as follows: ; Represents land parcels Belongs to the first The number of first-level POI categories, and the purity of the dominant POI category. and category intervals Adaptive determination of fusion weights To obtain the corrected probability distribution ; in, , These represent the first and second largest probability values ​​in the basic prediction, respectively; This represents the preset adaptive fusion weight determination function. In this invention, after the preset POI primary category and land use category are mapped and aligned, dimension K is equal to the total number of land use categories in set C. This indicates that the land parcels have been aligned by category. i It belongs to the candidate land use category c POI ratio; Steps 4-5) Based on the land parcel i DT neighborhood node set Calculate the neighborhood average probability distribution Neighborhood Consistency Coefficient Adaptive determination of smoothing weights Finally, the inference prediction probability distribution is obtained by fusion. ; in, Indicates land parcel i The set of DT neighborhood nodes Represents a set The number of neighboring nodes included in it; This represents the preset adaptive smoothing weight determination function. This indicates a normalization operation. Representing neighboring nodes j The predicted probability vector corrected by POI information; This leads to the final enhanced reasoning result. and final prediction confidence level ;in, This represents the probability distribution of the inference prediction obtained from the final fusion. Middle plot i Belongs to candidate land use category c The probability value.

[0014] In step 4-1), the node feature vector is specifically constructed as follows: ; in, Represents land parcels Node feature vectors; Indicates the proportional characteristics of each primary building category; This indicates the total number of buildings within a land parcel. Indicates the category richness characteristics within a land parcel; T This represents the transpose operation of a vector; In step 4-2), the satisfy: ; in, Represents land parcel nodes i With nodes j Are there any spatial connecting edges between them? and Representing land parcels and land plots The collection of buildings included; In step 4-3), the output basic prediction probability distribution is: ;in, and These represent the weight parameters and bias parameters of the output layer, respectively. Represents a node i The hidden layer feature representation vector obtained after weighted aggregation of neighborhood features; In step 4-4), the corrected probability distribution is obtained as follows: ; in, Represents land parcels The predicted probability vector after correction by POI information.

[0015] In steps 4-5), the final fusion-derived inference prediction probability distribution takes the following form: .

[0016] In step 4, constructing a multi-scale urban building knowledge graph specifically includes the following sub-steps: Steps 4-6) Construct Delaunay triangulations based on Landuse units with inferred land use categories to capture their potential spatial relationships; Steps 4-7) Combine road network information to prune the relationship edges between the functional area units of the block, retain the relationship edges with stronger spatial correlation and semantic functional relevance, and implement high-level clustering and merging to construct a higher-level spatial expression structure and form a region knowledge graph. Steps 4-8) By establishing multi-level parent-child relationships among multi-granular spatial unit nodes such as Building, Landuse, and Region, a multi-level urban building geographic knowledge graph is constructed and its associated expression and storage are completed.

[0017] Compared with the prior art, the present invention has the following technical effects: This invention overcomes the problems of existing land use mapping methods, such as the lack of potential relationships between blocks, insufficient semantic association characterization, and weak multi-scale expression ability, caused by over-reliance on road buffer zone division, by constructing a multi-level association expression framework of "building-block functional area-regional block" based on geographic knowledge graph. At the same time, in view of the problem that existing category reasoning methods rely heavily on manually labeled samples and do not make full use of surrounding crowdsource information, this invention proposes a land use category reasoning method that combines rule-based pruning clustering and graph attention network. Thus, under the conditions of limited labeled samples, high degree of urban function mixing, and heterogeneous multi-source information, this invention improves the accuracy of land use function identification, cross-scale semantic reasoning ability, and interpretability of multi-scale mapping results. Attached Figure Description

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is an overall framework diagram of the present invention; Figure 3 This is a schematic diagram of the multi-layer knowledge graph construction result according to an embodiment of the present invention. Detailed Implementation

[0019] like Figure 1 As shown, a method for constructing and aggregating building knowledge graphs for multi-scale representation includes the following steps: Step 1: Data Acquisition and Preprocessing; Collect building vector map datasets, perform data cleaning, geometric correction, and attribute completion operations on the collected building data to form a standardized basic building dataset; Step 2: Modeling the spatial and semantic relationships of buildings; Based on building data, a Delaunay triangulation is constructed to capture the spatial proximity and topological associations between buildings. At the same time, by combining multi-source information such as building function semantics, semantic similarity relationships between buildings are constructed to characterize the potential associations of buildings in terms of function type, name attributes and local spatial range, providing spatial and semantic basis for subsequent multi-scale merging. Step 3: Single-story building aggregation and knowledge graph construction; Based on the building relationship network constructed in step 2, graph theory methods are used to achieve building clustering and merging. During the merging process, land use information is used as the core semantic constraint, and spatial proximity and semantic similarity are combined to hierarchically aggregate buildings and form semantically consistent building group units. Based on this, an initial single-layer building knowledge graph is constructed with buildings as nodes, and spatial relationships and semantic associations are used as edges to form a basic graph structure of "buildings" at the single-layer level. Step 4: Construction of a multi-scale building knowledge graph; Based on the knowledge graph constructed in step 3, rule integration and reasoning are used, combined with spatial division information such as road networks, land parcel boundaries, and administrative divisions, to further aggregate and organize building cluster units and reason about land use categories, forming higher-level spatial units such as "building cluster - block - area block". By establishing parent-child relationships between different levels, a multi-scale urban building knowledge graph is constructed, realizing the structured expression and associated storage of urban space at multiple granularities such as building level, building cluster level, and area level, supporting multi-scale query, analysis, and mapping applications.

[0020] In step 1, building vector map data is collected, and the building dataset undergoes data preprocessing and standardization verification. The building vector map dataset is preferably selected from urban scenes with dense building distribution, well-developed road networks, and interwoven urban functions such as commerce, residence, public services, and industry. This ensures that the data simultaneously contains rich building semantic information and complex spatial adjacency relationships, thus providing a data foundation for building relationship modeling, regional aggregation, and multi-level knowledge graph construction. Furthermore, to enhance the applicability of the proposed method under different geographical conditions, the dataset can cover multiple regional scenes with differentiated spatial structural characteristics. Preprocessing of the raw data includes data cleaning, geometric correction, attribute completion, semantic enhancement, and category system alignment to ensure the correctness, consistency, and stability of building spatial location, geometric shape, and attribute information, ultimately forming a standardized basic building dataset.

[0021] In step 2, the specific steps are as follows: Based on the standardized building data obtained in Step 1, a hierarchical entity representation structure for buildings is first constructed. Specifically, building names, functional categories, center point coordinates, and attribute information such as offset direction and aspect ratio obtained during the preprocessing stage are bound to the corresponding building nodes to form a unified graph structure representation. This method simplifies subsequent graph calculations and relationship construction while preserving the spatial location features, semantic attributes, and morphological descriptions of building entities relatively completely, providing a foundation for subsequent building clustering and cross-level node generation.

[0022] After abstracting the building nodes, a Delaunay triangulation is constructed based on the building center points to characterize the spatial relationships between buildings. Building pairs with direct connections within the triangulation are defined as candidate spatial adjacency relationships. Further, road network information is incorporated to subdivide these candidate spatial adjacency relationships into general adjacency relationships and cross-road adjacency relationships, representing different strengths of spatial association between buildings under road obstruction conditions. Simultaneously, to construct semantic similarity relationships between buildings, a fuzzy string matching method is used to calculate building name similarity. This is combined with the reclassified building function category system to comprehensively determine the functional consistency and name attribute association between buildings. Spatial distance constraints are also set to retain building pairs with strong associations and functional consistency. Finally, building nodes and their spatial and semantic relationship edges are stored together in a graph database to form a building-layer knowledge graph, the graph structure of which can be represented as follows: ; in, Represents the set of entity nodes in a building layer. This represents the set of edges representing relationships between entities.

[0023] The specific steps for the rule-based pruning clustering operation mentioned in step 3 are as follows: Based on the building layer knowledge graph constructed in step 2, rule-based pruning and clustering operations are performed. Given the differences in spatial meaning and semantic strength corresponding to different relation sources, directly using all relations for subsequent clustering can easily introduce redundant connections and weakly related edges, thus affecting the accuracy of building group identification. Therefore, this invention designs relation pruning rules based on query statements in the graph database to remove weakly related relation edges and retain entity connections with high spatial and semantic consistency.

[0024] Specifically, in the first pruning step, building pairs that simultaneously possess cross-road adjacency and semantic similarity relationships are processed. Considering the spatial barrier effect of roads on building aggregation, when the semantic similarity between building pairs is lower than a preset threshold, it is considered that they lack sufficient aggregation basis at both the spatial and semantic levels, and therefore the corresponding relationship is deleted; conversely, when the semantic similarity is not lower than the threshold, the cross-road adjacency relationship is retained, and the semantic similarity relationship is deleted to avoid repeatedly expressing the association between the same building pair, which satisfies: ; in, and These represent the numbers of any two building nodes; Represents buildings With buildings Whether there is a cross-road relationship between them, a value of 1 indicates that it exists, and a value of 0 indicates that it does not exist; Represents buildings With buildings Does a semantic similarity exist between them? Represents buildings With buildings semantic similarity; and These represent the preserved states of the corresponding relationships after the first pruning step; In the second pruning step, building pairs that simultaneously possess both general adjacency and semantic similarity relationships are processed. To reduce relationship redundancy and highlight spatial continuity, when buildings have both adjacency and semantic similarity relationships, adjacency relationships are retained first, while semantic similarity relationships are deleted, satisfying the following condition: ; in, Represents buildings With buildings Whether there is an adjacency relationship between them, a value of 1 indicates that the two are directly adjacent in space; Indicate whether there is a semantic similarity relationship between the two; and These represent the preserved states of adjacency relations and semantic similarity relations after the second pruning step, respectively. In the third pruning step, to avoid retaining building connections with excessively different forms solely based on topological proximity or textual semantics, building form similarity is introduced as a further constraint. This form similarity is weighted by building area similarity, aspect ratio similarity, and orientation angle similarity, and its calculation formula is as follows: ; in, , and The similarity is represented by area, aspect ratio, and orientation angle, respectively, all calculated as the ratio of the minimum to the maximum value of the corresponding indicators for the two buildings, with values ​​ranging from [0,1]. Considering that building orientation has a stronger representation effect on the consistency of group morphology, this paper assigns a higher weight to orientation angle similarity. Finally, only relationship edges with morphological similarity of not less than 0.6 are retained, and relationships below this threshold are removed.

[0025] After pruning according to the above rules, buildings are clustered and merged based on the retained relation edges. Specifically, the connected component identification method in graph theory is used to aggregate building nodes in the relation network, forming building group units with strong spatial-semantic consistency. Furthermore, using land use information as the core semantic constraint, hierarchical aggregation of building groups is implemented, making the resulting groups more consistent in functional organization. On this basis, an initial single-layer knowledge graph is constructed with buildings as nodes and spatial relations and semantic associations as edges, forming the basic graph structure of the "building" hierarchy.

[0026] Subsequently, during the aggregation process of building-block functional areas, a merging operation is further implemented from the building level to the land use level. This process includes two key steps: First, based on the spatial location matching relationship between the building clustering results and the Areas of Interest (AOIs) formed by road expansion, an inclusion relationship is established between the buildings and the initial spatial units of the AOIs. Based on the entity associations retained after rule pruning, multiple AOIs in the same connected component are merged into the same block functional area unit, Landuse. Second, after the Landuse unit is formed, land use category inference is performed on the Landuse unit by combining the internal building composition characteristics and the external crowdsourced environment characteristics.

[0027] In the construction of Landuse unit-level features, the information affecting land use category determination is divided into two parts: internal representation and external constraint information. The internal representation mainly reflects the functional composition of buildings within the Landuse unit; the external constraint information mainly reflects the spatial consistency reflected by the semantic distribution of POIs within the unit and the DT adjacency relationships between units.

[0028] Specifically, firstly, based on the category mapping relationship established in the preprocessing stage, the secondary functional categories of buildings and POIs are uniformly mapped to primary categories. Then, for each Landuse unit, the proportion of each primary category corresponding to its internal buildings is statistically analyzed, and indicators such as the total number of buildings and category richness are supplemented to form the internal functional feature vector of that unit. Furthermore, initial pseudo-labels can be generated using a majority voting method based on the frequency of building categories within the unit, serving as supervision signals during the training stage of the graph attention network.

[0029] In the design of the category reasoning model, a graph attention network based on building composition features is adopted as the basic reasoning model. The model uses Landuse units as graph nodes, the building composition feature vectors within the unit as node attributes, and constructs graph connection edges based on the shared building relationships between units. Through the graph attention mechanism, information propagation and category determination between adjacent nodes are realized, and the basic probability distribution of each Landuse unit on the candidate land use category is output.

[0030] In step 3, the category reasoning module uses the mKGM method based on the graph attention network GAT to perform category reasoning on the land parcel nodes, specifically including the following steps: Step S1: Construct the architectural composition features and pseudo-labels within the Landuse unit; Landuse units corresponding to the functional zones of a block are used as graph nodes. First, based on the category mapping relationship established in the preprocessing stage, the original secondary functional categories of buildings within the unit are uniformly mapped to primary categories. For any land parcel... The number of each building category within the node is counted, and a node feature vector is constructed consisting of category proportion features, total number of buildings features, and category richness features. ; in, Represents land parcels Node feature vectors; Indicates the proportional characteristics of each primary building category; This indicates the total number of buildings within a land parcel. This indicates the richness of categories within a land parcel; K represents the total number of primary categories.

[0031] The proportional characteristics of each primary building category can be expressed as follows: ; in, Represents land parcels Belongs to the first The number of buildings in each of the first-level building categories.

[0032] The total number of buildings can be represented as: ; Category richness features can be represented as: ; in, This represents an indicator function, which takes the value 1 when the condition within the parentheses is true, and 0 otherwise.

[0033] Based on this, land parcels can also be generated using majority voting, based on the frequency of building categories within a unit. Initial pseudo-tags: ; in, Represents land parcels The initial pseudo-tag; Represents the set of all candidate land use categories; Represents land parcels Belongs to category The number of buildings.

[0034] Correspondingly, the confidence level of this initial pseudo-label can be expressed as: ; in, Represents land parcels The pseudo-label confidence score is used to characterize the consistency of building categories within the land parcel.

[0035] Step S2: Construct a functional area relationship diagram of the block based on shared building relationships; Establish connections between nodes based on whether functional zones within a block share buildings. When two functional zones contain at least one identical building, establish a connection edge between the corresponding nodes to form a functional zone association graph, whose adjacency relationships satisfy: ; in, Represents land block nodes With nodes Are there any connecting edges between them? and Representing land parcels and land plots The collection of buildings included; Step S3: Input the node feature matrix and the street functional area association graph into the graph attention network to obtain the basic inference result mKGM*; The node feature matrix obtained in step S1 and the street functional area association map obtained in step S2 are jointly input into the graph attention network. Through linear transformation and neighborhood attention allocation, the features of adjacent nodes are weighted and aggregated. The formula for calculating the unnormalized attention coefficient is as follows: ; Where, represents a node For nodes The unnormalized attention coefficient; Represents the attention parameter vector; Represents a linear transformation matrix; and Representing nodes respectively and nodes The input feature vector; This indicates a vector concatenation operation.

[0036] The normalized formula for attention weights is: ; in, Represents a node Assigning nodes when aggregating neighborhood information Attention weights; Represents a node The set of neighboring nodes; After attention weighting, the hidden representation of a node can be represented as: ; in, Represents a node The hidden representation; This represents a non-linear activation function.

[0037] The hidden representation obtained through attention aggregation is input to the output layer to obtain the basic probability distribution of each block's functional area belonging to each candidate category: ; in, Represents land parcels The basic predicted probability vector across all candidate categories; and These represent the weight parameters and bias parameters of the output layer, respectively.

[0038] The category with the highest probability is taken as the result of the basic category inference, that is: ; in, This represents the basic land use inference result obtained solely based on the internal building composition characteristics and shared building map structure, i.e., mKGM*; Represents land parcels Category The basic predicted probability.

[0039] The corresponding baseline prediction confidence level is: ; Step S4: Introduce the POI category distribution within the unit and perform posterior soft fusion on the basic prediction probabilities; After obtaining the basic category inference results, the basic predicted probabilities are further corrected posteriorly using the category composition of POIs within the unit. For land parcels... Count the number of first-level categories for each POI within it, and construct the corresponding category ratio vector: ; in, Represents land parcels POI category proportion vector; Represents land parcels Inner The proportion of POIs is defined as follows: ; in, Represents land parcels Belongs to the first Number of first-level POI categories.

[0040] Set up land plots The purity of the dominant class of POI is: ; in, Represents land parcels The proportion of the most important categories of internal POIs.

[0041] Let the difference between the highest probability value and the second highest probability value in the basic prediction results be: ; in, Represents land parcels The first most probable value in the basic prediction. Represents land parcels The second most probable value in the basic prediction.

[0042] According to the land parcel The POI fusion weights are adaptively determined based on the dominant category purity, number of effective POIs, baseline prediction confidence, and category margin. , can be represented as: ; in, This represents the adaptive determination function for the fusion weights; Represents land parcels The POI posterior fusion weights, and their value range satisfies: ; in, This indicates the preset maximum fusion weight.

[0043] The base predicted probabilities are weighted and fused with the prior proportions of POI categories to obtain the POI-corrected probability distribution: ; in, Represents land parcels The predicted probability vector corrected by POI information; This indicates a normalization operation.

[0044] Step S5: Smooth the category probabilities based on the DT adjacency relationship to obtain the enhanced inference result mKGM; After correction based on POI information, the predicted land parcel results are further smoothed for neighborhood consistency based on DT adjacency relationships. Let... Represents land parcels Given the set of DT neighborhood nodes, the average probability distribution of its neighborhood can be expressed as: ; in, Represents land parcels The neighborhood average probability vector; Represents land parcels The number of DT neighboring nodes.

[0045] Furthermore, land parcels are defined based on the consistency of the predicted categories among neighboring nodes. The neighborhood consistency coefficient is: ; in, Represents land parcels The proportion of the dominant category with the highest percentage in the neighborhood is used to reflect the degree of consistency of neighborhood categories.

[0046] Based on neighborhood consistency coefficient and land parcels The system adaptively determines neighborhood smoothing weights based on its own prediction confidence. , can be represented as: ; in, This represents the adaptive determination function for smooth weights; Represents land parcels The DT neighborhood smoothing weights.

[0047] Therefore, the probability corrected by POI is fused with the neighborhood average probability to obtain the final predicted probability distribution: ; in, Represents land parcels The final predicted probability vector after joint correction by POI and DT.

[0048] The category with the highest probability is taken as the final land use category, resulting in the enhanced inference result mKGM: ; in, Represents land parcels The final land use category inference result is the mKGM output.

[0049] The corresponding final prediction confidence level is: ; In step 4, based on the Building-Landuse two-layer knowledge graph that has been constructed in step 3, a higher-level spatial representation structure is further constructed. Specifically, a Delaunay triangulation is constructed based on Landuse units to capture potential relationships between functional areas of a block. Subsequently, road network information is used to prune the edges between functional areas, retaining those with stronger relationships, and Landuse-level clustering and merging are then implemented. Further, a Region-level knowledge graph is constructed for higher-level regional representations, achieving the construction of a multi-level knowledge graph of Building-Landuse-Region. By establishing parent-child relationships between nodes at different levels, structured representation and associated storage of multi-granularity urban spatial units at the building, building cluster, and region levels can be achieved, thus supporting multi-scale querying, multi-scale analysis, and multi-scale mapping applications.

[0050] The proposed method constructs and aggregates a building knowledge graph oriented towards multi-scale expression. It innovatively introduces crowdsourced information, semantic correlations between blocks of interest, and cross-scale semantic and geographical correlations, constructing a multi-level geographic knowledge graph. This addresses the problems of existing street function reasoning methods, such as the lack of potential inter-block relationships, insufficient capture of semantic correlations, and weak multi-scale expression capabilities due to over-reliance on road buffer zone divisions. Furthermore, addressing the issues of existing category reasoning methods' strong dependence on manually labeled data and insufficient utilization of surrounding crowdsourced information semantics, a land use category reasoning method combining rule-based pruning clustering and graph attention networks is proposed. This improves the accuracy of land use function identification, cross-scale semantic reasoning capabilities, and the interpretability of multi-scale mapping results, even under conditions of limited labeled samples, complex urban function mixtures, and heterogeneous multi-source information. This enables the structured organization and associated expression of urban spatial units at different granularities, such as building level, building cluster level, and region level, providing support for multi-scale querying, multi-scale analysis, and multi-scale mapping.

[0051] Example: 1) Experimental setup: This invention was implemented in a Windows operating system environment, using Python as the primary development language and combining it with the Neo4j graph database for storing, managing, and querying multi-level geographic knowledge graphs. The experimental software environment was built on an Anaconda virtual environment, using PyTorch 3.9 and Python version 3.9.21. For deep learning implementation, PyTorch 2.5.1 was used as the model development framework, and PyTorch Geometric 2.6.1 was used as the graph neural network computing library to construct, train, and predict the graph attention network in the street functional area category reasoning. Simultaneously, the Neo4j 5.28.1 Python driver was used to implement the connection and interaction between the Python program and the graph database, and data processing was performed using Pandas 1.5.3.

[0052] In terms of the graph database environment, this embodiment uses Neo4j Community 5.26.0 as the storage platform for the geographic knowledge graph. The constructed knowledge graph includes building nodes, block functional area nodes, road nodes, and region nodes, with corresponding node labels of Building, LandUse, Road, and Region, respectively. Regarding relational schemas, the knowledge graph includes adjacency relationships between buildings (ADJACENT), semantic similarity relationships between buildings (SEMANTIC_SIMILAR), cross-road adjacency relationships between buildings (ACROSS_ROAD), dependency relationships between buildings and block functional areas (Belong_To_BL), adjacency relationships between block functional areas (Landuse_ADJACENT), and dependency relationships between block functional areas and region blocks (Belong_To_LR). Through the definition of the above node labels and relational types, multi-level geographic knowledge representation between the building layer, block functional area layer, and region layer can be achieved.

[0053] 2) Analysis of experimental results: Table 1 presents the downstream land function inference results aggregated using this method. Overall, the method achieves good classification performance, with OA of 0.864, Precision of 0.885, Recall of 0.864, F1 score of 0.867, and Kappa score of 0.745. This indicates that the multi-layer knowledge graph can effectively integrate architectural semantics and spatial relationship information, providing stable support for land use identification.

[0054]

[0055] Regionally, Qiaokou District performed best, with OA and F1 scores of 0.872 and 0.867 respectively; Jiang'an District followed, with OA of 0.851 and F1 of 0.861. Jiang'an District was relatively lower, with OA of 0.851 and F1 of 0.861. There were slight differences in semantic rules between different regions, but the overall rules were similar. The data shows that the semantic consistency within the AOI has a direct impact on the downstream recognition performance.

[0056] This method, based on the overall experiment in Wuhan, further compares the performance of SVM, Random Forest, GCN, and TransE methods in identifying street functions based on the results of graph-based inference clustering. The results show that traditional machine learning methods generally perform poorly. Specifically, SVM has an OA of 80.06%, an F1 score of 80.03%, and a Kappa score of 62.12%. Random Forest shows a slight improvement, with an OA of 80.15%, an F1 score of 80.32%, and a Kappa score of 62.44%. In contrast, methods utilizing graph structure information perform significantly better. GCN achieves an OA of 81.44% and an F1 score of 81.16%, while TransE has an OA of 81.07% and an F1 score of 80.58%. This indicates that by introducing graph structures or knowledge representations, the models can better capture the spatial-semantic relationships between land use units.

[0057] Table 2 shows the comparison results of land use models on the dataset:

[0058] Among the various methods, our proposed method achieved the best results. In Mkgm* without multi-source data, OA, Precision, Recall, F1, and Kappa reached 81.72%, 83.81%, 81.72%, 82.22%, and 66.84%, respectively. After introducing multi-source data, OA, Precision, Recall, F1, and Kappa in Mkgm reached 82.64%, 84.41%, 82.64%, 82.94%, and 68.09%, respectively, all outperforming other comparative models. This indicates that the graph attention mechanism can more effectively utilize the differences in importance of neighboring nodes in the graph, thereby improving the accuracy of street function identification.

[0059]

[0060] Table 3 systematically defines the framework structure of the constructed geographic knowledge graph. This knowledge graph uses buildings, land use blocks, and regional blocks as core entity units, representing multi-level objects in urban space at the building, block, and regional levels, respectively. Simultaneously, it uses building proximity relationships, building-road-crossing proximity relationships, semantic similarity proximity relationships, land use proximity relationships, and hierarchical membership relationships as relational expressions to describe the geometric adjacency, road barriers, semantic associations, and inclusion / subordination features between different spatial units. Through this entity-relationship organization method, a hierarchical relational expression structure from buildings to land use blocks to regional blocks can be formed, providing a unified knowledge organization foundation for multi-scale spatial aggregation, category reasoning, and cartographic representation, and supporting subsequent spatial relationship retrieval, rule reasoning, and multi-level cartographic applications.

[0061] In summary, this invention proposes a method for constructing and aggregating building knowledge graphs for multi-scale representation, primarily designing a multi-level geographic knowledge graph construction process. This method, by introducing inter-block relationships and inter-layer semantic relationships combined with geographic connections, fully preserves feature semantic information, addressing the problems of existing street-level functional mapping methods that suffer from insufficient capture of potential inter-block relationships, inadequate semantic relevance, and weak multi-scale representation capabilities due to over-reliance on road buffer zones. Furthermore, it utilizes rule pruning to achieve associative clustering, merging strongly related buildings to form mid-level regions, i.e., street-level functional zones. Simultaneously, it combines GAT (Geographic Information Classification) to synthesize internal features and infer regional block categories. This invention verifies the effectiveness of the proposed method, significantly improving accuracy and semantic preservation capabilities in multi-scale mapping and land use function inference scenarios.

Claims

1. A method for constructing and aggregating building knowledge graphs for multi-scale representation, characterized in that, Includes the following steps: Step 1: Obtain building vector map data, perform data preprocessing, and form a standardized building basic dataset; Step 2: Based on the building dataset, perform spatial and semantic relationship modeling to construct the basic graph structure of building hierarchy; Step 3: Perform rule-based pruning and graph theory clustering on the constructed basic graph structure to form building group units, thereby constructing a single-story building knowledge graph; Step 4: Integrate spatial constraint information with multi-source geographical features to infer land use categories, establish connections between different levels, and construct a multi-scale urban building knowledge graph.

2. The method according to claim 1, characterized in that, Step 1 specifically includes the following sub-steps: Step 1-1) Collect building vector map data. The building vector map data is selected from regional scenes with high building density, complete road network structure and at least one urban functional space intertwined with commerce, residence, public services and industry. Steps 1-2) Perform data cleaning, geometric correction, attribute completion, semantic improvement, and category system alignment on the collected building vector map data to ensure that the buildings meet the requirements for subsequent processing in terms of spatial location, geometric integrity, and attribute information consistency, and finally form a standardized building basic dataset.

3. The method according to claim 2, characterized in that, Step 2 specifically includes the following sub-steps: Step 2-1) Abstract each building entity into a graph node and bind it with the building name, function category, center point coordinates and geometric description attributes extracted in the preprocessing stage to construct the entity representation structure of the building hierarchy; Step 2-2) Construct a Delaunay triangulation based on the building center point, capture the spatial proximity and topological association between buildings, and define the building pairs with direct connection relationships in the triangulation as candidate spatial adjacency relationships; Steps 2-3) Combining road network information, the candidate spatial adjacency relationships are further subdivided into general adjacency relationships and cross-road adjacency relationships; Steps 2-4) use fuzzy string matching to calculate building name similarity, and combine this with the reclassified building function category system to comprehensively determine the functional consistency and name attribute correlation between buildings, constructing semantic similarity relationships between buildings, forming a basic hierarchical graph structure of buildings, formally represented as: ； in, Represents the set of entity nodes in a building layer. This represents the set of edges representing relationships between entities.

4. The method according to claim 1, characterized in that, In step 3, performing rule-based pruning specifically includes the following sub-steps: Step 3-1) Perform the first step of rule pruning for building pairs that simultaneously have cross-road adjacency and semantic similarity relationships: when a building i With buildings j semantic similarity between When the value is below a preset first threshold of 0.9, its cross-road adjacency relationship is removed. The retained state after pruning is expressed as follows: ； in, i and j These represent the node numbers of any two buildings; Represents buildings i With buildings j Whether there is a cross-road relationship between them, a value of 1 indicates that it exists, and a value of 0 indicates that it does not exist; Represents buildings i With buildings j Whether there is a semantic similarity relationship between them, a value of 1 indicates that there is, and a value of 0 indicates that there is no; Represents buildings i With buildings j The semantic similarity between the function and the name; and These represent the preserved states of cross-road relationships and semantic similarity relationships after the first pruning step, respectively. Step 3-2) Perform the second step of rule pruning for building pairs that simultaneously have both general adjacency and semantic similarity relationships: When buildings have both adjacency and semantic similarity relationships, prioritize retaining the adjacency relationship and delete the semantic similarity relationship to reduce redundancy. The retained state is expressed as: ； in, Represents buildings i With buildings j Whether there exists a general adjacency relationship between them that is directly adjacent in space, a value of 1 indicates that they exist, and a value of 0 indicates that they do not exist. and These represent the preserved states of general adjacency relations and semantic similarity relations after the second pruning step; " indicates a pruning logic mapping operation; Step 3-3) Introduce building shape similarity as a constraint for the third step of rule pruning. The calculation formula is as follows: ； in, A comprehensive index representing the morphological similarity between buildings. , and These represent the similarity of the corresponding areas, aspect ratios, and orientation angles of the two buildings, respectively.

5. The method according to claim 1, characterized in that, In step 3, forming the building group unit specifically includes the following sub-steps: Step s3-1) Using the connected component identification method in graph theory, the building nodes corresponding to the relation edges retained after rule pruning are aggregated to form building group units with strong spatial-semantic consistency; Step s3-2) Based on the spatial location matching relationship between the building group clustering results and the AOI formed by road expansion, establish the inclusion relationship, and according to the real topological associations retained after rule pruning, merge multiple AOIs in the same connected component into the same block functional area unit, which is defined as the Landuse unit.

6. The method according to any one of claims 1 to 5, characterized in that, In step 4, land use category reasoning is performed on the block functional area unit Landuse, which specifically includes the following sub-steps: Step 4-1) Statistically analyze the proportion characteristics of each building at the first-level within the functional unit of the block. Characteristics of the total number of buildings and category richness features , constituting a plot of land i Node feature vectors ; Initial pseudo-labels are generated using a majority voting method. and its confidence level ,in Indicates land parcel Belongs to category The number of buildings; in, K This indicates the total number of preset primary building categories; k and c Index representing the category; Indicates land parcel i The number of buildings belonging to the kth first-level building category; This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. Step 4-2) Establish spatial adjacency constraints based on whether the plots share buildings, and construct a functional area association graph of the block. The spatial connection edges of its topological connection matrix are: ; Step 4-3) Input the node feature vectors and association graph into the network, and calculate the unnormalized attention coefficients. Normalized attention weights The hidden representation is obtained after feature aggregation. Output the basic prediction probability distribution ; in, The attention parameter weight vector is represented by W; the trainable linear transformation feature map matrix is ​​represented by ||; and the vector concatenation operator is represented by ||. Represents a node i The set of neighborhood nodes in the association graph Represents a node i Neighborhood nodes in the association graph j eigenvectors, Represents a non-linear activation function; This leads to the basic land use inference results. and prediction confidence ;in, c Indicates category index, C This represents the set of all candidate land use categories. Represents the basic prediction probability distribution The central plot belongs to the candidate land use category c The probability value; Step 4-4) Calculate the category proportion vector at the first level of POI. ,in Represents land parcels Inner The proportion of POIs is defined as follows: ; Represents land parcels Belongs to the first The number of first-level POI categories, and the purity of the dominant POI category. and category intervals Adaptive determination of fusion weights To obtain the corrected probability distribution ; in, , These represent the first and second largest probability values ​​in the basic prediction, respectively; This represents a preset adaptive fusion weight determination function. In this invention, after the preset POI primary category and land use category are mapped and aligned, dimension K equals the total number of land use categories in set C. This indicates that the land parcels have been aligned by category. i It belongs to the candidate land use category c POI ratio; Steps 4-5) Based on the land parcel i DT neighborhood node set Calculate the neighborhood average probability distribution Neighborhood Consistency Coefficient Adaptive determination of smoothing weights Finally, the inference prediction probability distribution is obtained by fusion. ; in, Indicates land parcel i The set of DT neighborhood nodes Represents a set The number of neighboring nodes included in it; This represents the preset adaptive smoothing weight determination function. This indicates a normalization operation. Representing neighboring nodes j The predicted probability vector corrected by POI information; This leads to the final enhanced reasoning result. and final prediction confidence level ;in, This represents the probability distribution of the inference prediction obtained from the final fusion. Middle plot i Belongs to candidate land use category c The probability value.

7. The method according to claim 6, characterized in that, In step 4-1), the node feature vector is specifically constructed as follows: ； in, Represents land parcels Node feature vectors; Indicates the proportional characteristics of each primary building category; This indicates the total number of buildings within a land parcel. Indicates the category richness characteristics within a land parcel; T This represents the transpose operation of a vector; In step 4-2), the satisfy: ； in, Represents land parcel nodes i With nodes j Are there any spatial connecting edges between them? and Representing land parcels and land plots The collection of buildings included; In step 4-3), the output basic prediction probability distribution is: ;in, and These represent the weight parameters and bias parameters of the output layer, respectively. Represents a node i The hidden layer feature representation vector is obtained after weighted aggregation of neighborhood features.

8. The method according to claim 6, characterized in that, In step 4-4), the corrected probability distribution is obtained as follows: ； in, Represents land parcels The predicted probability vector after correction by POI information.

9. The method according to claim 6, characterized in that, In steps 4-5), the final fusion-derived inference prediction probability distribution takes the following form: 。 10. The method according to claim 1, characterized in that, In step 4, constructing a multi-scale urban building knowledge graph specifically includes the following sub-steps: Steps 4-6) Construct Delaunay triangulations based on Landuse units with inferred land use categories to capture their potential spatial relationships; Steps 4-7) Combine road network information to prune the relationship edges between the functional area units of the block, retain the relationship edges with stronger spatial correlation and semantic functional relevance, and implement high-level clustering and merging to construct a higher-level spatial expression structure and form a region knowledge graph. Steps 4-8) By establishing multi-level parent-child relationships among multi-granular spatial unit nodes such as Building, Landuse, and Region, a multi-level urban building geographic knowledge graph is constructed and its associated expression and storage are completed.

Images