Building intelligent comprehensive selection method fusing multiple features

Abstract

The invention relates to the field of map synthesis and big data analysis, in particular to an intelligent comprehensive building selection method fusing multiple features. According to the method, shp file data of a target city building and original POI data of the target city building are processed to obtain a final graph structure and building characteristics of the building, geometric characteristics and semantic characteristics of the building are fused in the building characteristics, and then the building is classified firstly by using an idea of coarse clustering and fine selection; and building selection is carried out by using a graph convolutional neural network. According to the method provided by the invention, various attribute features of the building can be integrated, the multi-dimensional features of the building are comprehensively reflected in the selection process, and the flexibility and adaptability of the selection method are enhanced by using advanced technologies such as deep learning and a graph convolutional neural network; the method is suitable for building selection in different geographic areas and improves the accuracy of a selection result.

Description

Technical Field

[0001] The present invention relates to the field of map synthesis and big data analysis, and in particular to an intelligent comprehensive selection method for buildings integrating multiple features. Background Art

[0002] Map generalization is one of the core issues in cartography. In particular, when large-scale spatial data is reduced to small-scale spatial data, it is crucial to effectively abstract and summarize spatial data. Cartographic generalization is not only a technical issue, but also scientific and engineering, involving multiple aspects such as scale transformation, integration and fusion of spatial data, and data analysis and mining. In cartography, residential buildings are usually presented in the form of surface elements. However, in the process of scale transformation, visual conflicts may occur, such as overlap or overlap. These visual interferences not only affect the readability of the map, but also affect the user's sensory experience. Therefore, how to effectively select buildings during the scale transformation process, discard unimportant buildings and retain key buildings, has become an important challenge in cartographic generalization.

[0003] In recent years, the research on building selection has gradually increased, and scholars have tried to explore it from different angles. These studies usually consider the semantic characteristics of buildings, such as the importance of buildings, administrative levels, etc., and rely on specific selection ideas for sorting and selection. Some methods use rough set theory to simplify attribute knowledge to evaluate the selection of residential areas, and some use the principle of information entropy to sort and select them. With the deepening of research, subsequent selection work began to focus on the geometric attributes of buildings. Although some studies have tried to use Delaunay triangulation and other technologies to organize buildings into graph structures and calculate their geometric features, they have ignored the semantic characteristics and contextual relationships of the target buildings, which makes it difficult for the final selection results to fully reflect the actual situation.

[0004] In addition, the problem of automated selection has always been one of the main challenges in map generalization research. Early selection methods mainly relied on index evaluation and square root formulas. Although they can effectively complete the selection task, they lack generalization ability and cannot be applied to different geographical regions. How to use the rapidly developing deep learning knowledge to carry out more efficient comprehensive selection work is one of the research directions in the field of map generalization. Summary of the invention

[0005] In order to solve the problem of ignoring the semantic features and contextual relationships of target buildings in the prior art, the present invention provides an intelligent comprehensive selection method for buildings that integrates multiple features.

[0006] The present invention provides a method for intelligent comprehensive selection of buildings integrating multiple features, comprising:

[0007] Step S1: obtaining the initial graph structure of the buildings in the target city and the geometric features of each building based on the shp file data of the buildings in the target city;

[0008] Step S2: obtaining semantic features of each building based on the original POI data of the target city; wherein the semantic features include importance features and functional semantic features;

[0009] Step S3: performing weighted processing on the initial graph structure based on the spatial distance between buildings and the functional semantic features to obtain a final graph structure of buildings in the target city;

[0010] Step S4: Coarsely cluster all buildings in the target city based on their spatial distribution;

[0011] Step S5: selecting a class of building groups in the coarse clustering results, and marking the buildings of interest in the class with selection labels, so as to train the graph convolutional neural network model to obtain a building selection model; wherein the training data used includes the building features and the final graph structure of the buildings marked with the selection labels; the building features include the geometric features and the semantic features;

[0012] Step S6: using the building selection model to select buildings for the unselected building groups in the rough clustering results respectively, merging the selection results of each building group to obtain a comprehensive selection result of the entire target city map.

[0013] Furthermore, step S1 specifically includes:

[0014] Step S101: constructing a Delaunay triangulation network based on the shp file data of the target city buildings;

[0015] Step S102: performing skinning processing on the Delaunay triangulation to obtain the initial graph structure of the target city building;

[0016] Step S103: extracting geometric features of each building based on the shp file data of the target city buildings, wherein the geometric features include area, perimeter, shape factor, density and minimum circumscribed rectangle direction.

[0017] Furthermore, the Delaunay triangulation is subjected to skinning processing, which specifically includes:

[0018] Step S1021: constructing a convex hull based on the Delaunay triangulation network and extracting edge points; the edge points are points located on the edge of the convex hull;

[0019] Step S1022: Calculate the area and shape factor of each triangle in the Delaunay triangulation, where the calculation formula of the shape factor S is as follows:

[0020]

[0021] Among them, A i represents the area of ​​the triangle, P i represents the perimeter of a triangle;

[0022] Step S1023: deleting triangles in the Delaunay triangulation based on constraint conditions; the constraint conditions include the number of triangle vertices belonging to the edge points, the area and shape factor of the triangle.

[0023] Furthermore, step S2 specifically includes:

[0024] Step S201: reclassifying the original POI data of the target city;

[0025] Step S202: sorting the reclassified POI data based on public recognition to obtain importance characteristics;

[0026] Step S203: using a greedy algorithm to determine the spatial relationship between the reclassified POI data to obtain a POI corpus;

[0027] Step S204: performing word vector conversion based on the POI corpus to obtain functional semantic features.

[0028] Furthermore, the step S202 specifically includes:

[0029] The reclassified POI data are sorted based on public recognition, different levels of weights are assigned to the POI data, and the weights are mapped to the buildings to obtain the importance characteristics of the buildings.

[0030] Furthermore, the step S203 specifically includes:

[0031] Step S2031: Initialize an optimal path set and calculate the path weight between any two POI data;

[0032] Step S2032: randomly selecting a POI data from all POI data as the starting point and adding it to the optimal path set, marking it as visited;

[0033] Step S2033: starting from the starting point, and selecting the POI data closest to the current point of interest and not visited as the next point of interest according to the path weight, adding the next point of interest to the optimal path set and marking it as visited;

[0034] Step S2034: repeat the above step S2033 until all POI data have been visited;

[0035] Step S2035: Obtain the POI data sequence according to the optimal path set, record the names of each POI data in sequence, and construct a POI corpus.

[0036] Furthermore, the step S204 specifically includes:

[0037] The POI corpus is used to train on a Word2Vec model based on CBOW to obtain vectorized representations of POI data of various categories, and the textual POI data is converted into digital data recognizable by a computer to obtain functional semantic features.

[0038] Furthermore, the step S3 specifically includes:

[0039] Step S301: performing distance weighting on the Delaunay triangulation based on the spatial distance between buildings to obtain a distance weight matrix;

[0040] Step S302: obtaining a function similarity weight matrix based on the vector cosine similarity relationship between the function semantic features;

[0041] Step S303: performing weighted processing on the initial graph structure based on the distance weight matrix and the functional similarity weight matrix to obtain a final graph structure.

[0042] Furthermore, the coarse clustering in step S4 specifically includes:

[0043] Step S401: Initialize K cluster centers based on the coordinates of the geometric center point of the building;

[0044] Step S402: iterative calculation, assigning each building to the cluster center closest to it, and updating the location of the cluster center;

[0045] Step S403: using the improved elbow method to determine the K value, by plotting the clustering error sum of squares curves corresponding to different K values, observing the inflection point, and selecting the K value that can retain the regional distribution characteristics to the maximum extent, thereby optimizing the clustering results;

[0046] Step S404: According to the K value, the original buildings are divided into K categories according to their spatial distribution.

[0047] Furthermore, the step S5 further includes:

[0048] During the training process, edges in the graph are randomly deleted to force the model to learn more robust feature representations, thereby improving the generalization ability of the model.

[0049] Beneficial effects of the present invention:

[0050] Based on the traditional building selection method, the present invention makes the selection process more scientific, automated and efficient through a number of key improvements, providing a robust solution for mapping work, with high market application value and broad promotion prospects.

[0051] (1) The present invention enhances the ability to describe buildings by comprehensively considering the geometric and semantic features of buildings, and also improves the scientificity and rationality of the selection results. The selection results of the present invention are more scientific and reasonable, and can more accurately reflect the importance of buildings in residential areas. This multi-dimensional feature fusion method significantly improves the quality of selection and provides users with more practical information.

[0052] (2) The present invention proposes a mapping idea of ​​"coarse clustering followed by fine selection", combined with a semi-supervised deep learning model, which can achieve efficient and automated building selection while reducing manual intervention. This improvement greatly saves time and labor costs, and provides a feasible solution for large-scale mapping work.

[0053] (3) Traditional methods often require a lot of manual annotation, while the semi-supervised learning method adopted by the present invention only requires a small amount of manual annotation to generate rich selection results. This feature not only improves work efficiency, but also makes the technology more feasible and flexible in practical applications.

[0054] (4) Through the use of graph convolutional neural networks, the model can deeply learn the complex spatial relationships between buildings, so that the selection process is not only based on a single feature, but also integrates multiple information. This rich spatial relationship learning provides a more comprehensive perspective for selection and helps to make more accurate judgments.

[0055] (5) The present invention introduces the DropEdge concept and forces the model to learn a more robust feature representation, thereby reducing the reliance on specific edge structures and solving the problem of the generalization of graph convolutional neural networks. This improves the model's adaptability to different data features and makes the selection results more stable in various geographical environments. This means that the selection method of the present invention can adapt to a variety of application scenarios, whether it is urban, rural or different types of residential areas, and can maintain high accuracy and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 One of the flow charts of a method for intelligent comprehensive selection of buildings integrating multiple features provided by an embodiment of the present invention;

[0057] Figure 2 A second flow chart of a method for intelligent comprehensive selection of buildings integrating multiple features provided by an embodiment of the present invention;

[0058] Figure 3 A schematic diagram of a process for constructing a POI corpus provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] like Figure 1 and Figure 2 As shown, an embodiment of the present invention provides a method for intelligent comprehensive selection of buildings integrating multiple features, including:

[0061] Step S1: obtaining the initial graph structure of the buildings in the target city and the geometric features of each building based on the shp file data of the buildings in the target city;

[0062] Step S101: constructing a Delaunay triangulation network based on the shp file data of the target city buildings;

[0063] Specifically, the geometric center points of each building in the shp file data of the target city buildings are first regarded as nodes of the graph, and then the edges in the undirected graph are constructed to obtain the Delaunay triangulation.

[0064] It should be noted that the shp file is a commonly used GIS data format, which is used to store and edit various geographic information data, including points, lines, surfaces, etc. The adjacent relationship between buildings can be obtained through the Delaunay triangulation, which can ensure that the spatial relationship between buildings is reasonably represented.

[0065] It can be understood that the method for constructing a Delaunay triangulation provided in the embodiment of the present invention is a preferred method, and there are other methods for constructing a Delaunay triangulation in the art which are not described here.

[0066] Step S102: peeling the Delaunay triangulation to obtain the initial graph structure of the target city buildings;

[0067] Step S103: extracting geometric features of each building based on the shp file data of the target city buildings, the geometric features including area, perimeter, shape factor, density and minimum circumscribed rectangle direction;

[0068] Step S2: obtaining semantic features of each building based on the original POI data of the target city; wherein the semantic features include importance features and functional semantic features, specifically including:

[0069] Step S3: weighting the initial graph structure based on the spatial distance and functional semantic features between buildings to obtain the final graph structure of the target city buildings;

[0070] Step S4: performing coarse clustering on all buildings in the target city based on the spatial distribution of the buildings; the spatial distribution is obtained according to the initial graph structure;

[0071] Step S5: selecting a class of building groups in the coarse clustering results, and marking the buildings of interest in the class with selected labels, so as to train the graph convolutional neural network model and obtain a building selection model; wherein the training data used includes its labeled data, building features of the buildings marked with the selected labels, and the final graph structure; the building features include geometric features and semantic features;

[0072] Specifically, according to the rough clustering of the differences in the spatial distribution of buildings, a group of buildings in K categories is selected for labeling. Before model training, the buildings of interest in this category are manually selected for labeling, which are used as the true values ​​for training and adjusting model parameters; the graph convolutional neural network model is used as the basic model for training, and the training data used includes its labeled data, the building features of the buildings labeled with the selected labels, and the final graph structure. The spatial relationship and features between the buildings are extracted through the multi-layer convolutional layers in the model to select the model. As a preferred embodiment, the trained model can also be evaluated, and the model performance can be optimized according to indicators such as accuracy, recall rate and F1-score to ensure the effectiveness of the final model in practical applications.

[0073] Furthermore, during the training process, edges in the graph are randomly deleted to force the model to learn more robust feature representations, thereby improving the generalization ability of the model.

[0074] It can be understood that in order to improve the overfitting phenomenon of the traditional GCN model, the DropEdge idea is adopted to randomly delete the edges in the graph during the training process to force the model to learn more robust feature representations, thereby improving the generalization ability of the model.

[0075] Step S6: using the building selection model to select buildings for the unselected building groups in the rough clustering results respectively, merging the selection results of each building group to obtain a comprehensive selection result of the entire target city map.

[0076] Specifically, the trained model is used to select the remaining unselected building groups separately to obtain whether the buildings in each building group are selected and the probability of selection and rejection. Then the selection results of each building group are merged to obtain the final comprehensive selection result of the entire city map.

[0077] The method provided by the embodiment of the present invention uses the shp file data and the original POI data of the target city building to obtain the geometric features, importance features and functional semantic features of the building, and based on this, performs coarse clustering and fine selection on the building to obtain the selection results of the target city building. It overcomes the problem of ignoring the semantic features of the target building and the relationship between the buildings in the prior art, and uses the graph structure, intelligent data conversion, accurate clustering method and semi-supervised learning technology to greatly improve the selection efficiency and accuracy of the building, providing an innovative and effective solution for mapping work, and has good market prospects and application value.

[0078] In one embodiment, the skinning process of the Delaunay triangulation network specifically includes:

[0079] Step S1021: construct a convex hull based on the Delaunay triangulation network and extract edge points; wherein the edge points are points located on the edge of the convex hull;

[0080] Step S1022: Calculate the area and shape factor of each triangle in the Delaunay triangulation, where the calculation formula of the shape factor S is as follows:

[0081]

[0082] Among them, A i represents the area of ​​the triangle, P i represents the perimeter of a triangle;

[0083] Step S1023: deleting triangles in the Delaunay triangulation network based on constraint conditions; the constraint conditions include the number of triangle vertices that are edge points, the area of ​​the triangle, and the shape factor.

[0084] It can be understood that by iteratively skinning the triangulated network using the constraint conditions described in this embodiment, the outermost narrow triangles can be removed without causing internal holes, thereby further optimizing the structure of the building diagram represented by the Delaunay triangulated network.

[0085] In one embodiment, step 2 specifically includes:

[0086] Step S201: reclassifying the original POI data of the target city;

[0087] Specifically, the reclassification classifies buildings based on the Standards for Urban Land Classification and Planning and Construction Land (GB 50137-2011).

[0088] Step S202: sorting the reclassified POI data based on public recognition to obtain importance characteristics;

[0089] Specifically, the reclassified POI data are sorted based on public awareness, different levels of weights are assigned to each type of POI data, and the weights are mapped to buildings to obtain the importance characteristics of each building.

[0090] Step S203: using a greedy algorithm to determine the spatial relationship between the reclassified POI data and construct a POI corpus;

[0091] like Figure 3 As shown in the figure, the process of building a POI corpus is as follows:

[0092] Step S2031: Initialize an optimal path set and calculate the path weight between any two POI data;

[0093] Specifically, the Euclidean distance is used to calculate the path weight between POI data; the Euclidean distance calculation formula is as follows:

[0094]

[0095] Among them, (x i ,y j ) and (x j ,y j ) represent the geographic horizontal and vertical coordinates of building i and building j in the POI data respectively.

[0096] It can be understood that the path weight is inversely proportional to the Euclidean distance, and the path weight can be used in the subsequent greedy selection process.

[0097] Step S2032: randomly selecting a POI data from all POI data as the starting point and adding it to the optimal path set, marking it as visited;

[0098] Step S2033: starting from the starting point, and selecting the POI data closest to the current point of interest and not visited as the next point of interest according to the path weight, adding the next point of interest to the optimal path set and marking it as visited;

[0099] Step S2034: repeat the above step S2033 until all POI data have been visited;

[0100] Step S2035: Obtain the POI data sequence according to the optimal path set, record the names of each POI data in sequence, and construct a POI corpus.

[0101] It can be understood that the corpus obtained through this step is to connect the separate POI data in series, and furthermore, to implicitly embed the spatial relationship in the corpus through the distance attribute.

[0102] It should be noted that the method for constructing a POI corpus provided in the embodiment of the present invention is a preferred method. There are other methods for constructing a POI corpus in the prior art that can also be applied to the present invention, which will not be elaborated in detail here.

[0103] Step S204: Perform word vector conversion based on the POI corpus to obtain functional semantic features.

[0104] Specifically, the POI corpus is used to train the Word2Vec model based on CBOW to obtain the vectorized representation of POIs of each category, and the textual POI data is converted into computer-recognizable digital data to obtain functional semantic features.

[0105] In one embodiment, step 3: weighting the initial graph structure based on the spatial distance and functional semantic features between buildings to obtain the final graph structure of the target city buildings, specifically includes:

[0106] Step S301: Perform distance weighting on the Delaunay triangulation based on the spatial distance between buildings to obtain a distance weight matrix W d ;

[0107] Step S302: Based on the vector cosine similarity relationship between functional semantic features, obtain the functional similarity weight matrix W f ; Among them, the cosine similarity relationship calculation formula is as follows:

[0108]

[0109] Among them, F i represents the functional semantic vectorization expression of building i, F j represents the functional semantic vectorized expression of building j, and F_sim(i,j) represents the similarity relationship between building i and building j.

[0110] Step S303: weighting the initial graph structure based on the distance weight matrix and the function similarity weight matrix to obtain the final graph structure; the weighting process can be expressed by the following formula:

[0111]

[0112] Among them, A wrepresents the final graph structure, A represents the initial graph structure, and W d represents the distance weight matrix, W f represents the functional similarity weight matrix, E represents the identity matrix, and represents the self-connection of buildings, with 1 for connection and 0 for no connection.

[0113] In one embodiment, the buildings may be roughly clustered using a K-means clustering method, specifically including:

[0114] Step S401: Initialize K cluster centers based on the coordinates of the geometric center point of the building;

[0115] Step S402: iterative calculation, assigning each building to the cluster center closest to it, and updating the location of the cluster center;

[0116] Step S403: using the improved elbow method to determine the K value, by plotting the clustering error sum of squares curves corresponding to different K values, observing the inflection point, and selecting the K value that can retain the regional distribution characteristics to the maximum extent, thereby optimizing the clustering results;

[0117] Step S404: According to the K value, the original buildings are divided into K categories according to their spatial distribution.

[0118] It can be understood that by coarse clustering, the original buildings are distributed into K categories. In the subsequent work, each type of building can be selected separately and then merged. This can better understand and capture the global geometric characteristics of the building data. It is also a new idea and method for solving complex graph data problems.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent comprehensive selection of buildings integrating multiple features, characterized in that: include: Step S1: obtaining the initial graph structure of the buildings in the target city and the geometric features of each building based on the shp file data of the buildings in the target city; Step S2: obtaining semantic features of each building based on the original POI data of the target city; wherein the semantic features include importance features and functional semantic features; Step S3: weighting the initial graph structure based on the spatial distance between buildings and the functional semantic features to obtain a final graph structure of buildings in the target city; Step S4: Coarsely cluster all buildings in the target city based on their spatial distribution; Step S5: selecting a class of building groups in the coarse clustering results, and marking the buildings of interest in the class with selection labels, so as to train the graph convolutional neural network model to obtain a building selection model; wherein the training data used includes the building features and the final graph structure of the buildings marked with the selection labels; the building features include the geometric features and the semantic features; Step S6: using the building selection model to select buildings for the unselected building groups in the rough clustering results respectively, merging the selection results of each building group to obtain a comprehensive selection result of the entire target city map.

2. The intelligent comprehensive selection method of buildings integrating multiple features according to claim 1 is characterized in that: Step S1 specifically includes: Step S101: constructing a Delaunay triangulation network based on the shp file data of the target city buildings; Step S102: performing skinning processing on the Delaunay triangulation to obtain the initial graph structure of the target city building; Step S103: extracting geometric features of each building based on the shp file data of the target city buildings, wherein the geometric features include area, perimeter, shape factor, density and minimum circumscribed rectangle direction.

3. The intelligent comprehensive selection method of buildings integrating multiple features according to claim 2 is characterized in that: The Delaunay triangulation is subjected to skinning processing, specifically comprising: Step S1021: constructing a convex hull based on the Delaunay triangulation network and extracting edge points; the edge points are points located on the edge of the convex hull; Step S1022: Calculate the area and shape factor of each triangle in the Delaunay triangulation, where the calculation formula of the shape factor S is as follows: Among them, A i represents the area of ​​the triangle, P i represents the perimeter of a triangle; Step S1023: deleting triangles in the Delaunay triangulation based on constraint conditions; the constraint conditions include the number of triangle vertices belonging to the edge points, the area and shape factor of the triangle.

4. The intelligent comprehensive selection method of buildings integrating multiple features according to claim 1 is characterized in that: Step S2 specifically includes: Step S201: reclassifying the original POI data of the target city; Step S202: sorting the reclassified POI data based on public recognition to obtain importance characteristics; Step S203: using a greedy algorithm to determine the spatial relationship between the reclassified POI data to obtain a POI corpus; Step S204: performing word vector conversion based on the POI corpus to obtain functional semantic features.

5. The intelligent comprehensive selection method of buildings integrating multiple features according to claim 4 is characterized in that: The step S202 specifically includes: The reclassified POI data are sorted based on public recognition, different levels of weights are assigned to the POI data, and the weights are mapped to the buildings to obtain the importance characteristics of the buildings.

6. The intelligent comprehensive selection method of buildings integrating multiple features according to claim 4 is characterized in that: The step S203 specifically includes: Step S2031: Initialize an optimal path set and calculate the path weight between any two POI data; Step S2032: randomly selecting a POI data from all POI data as the starting point and adding it to the optimal path set, marking it as visited; Step S2033: starting from the starting point, and selecting the POI data closest to the current point of interest and not visited as the next point of interest according to the path weight, adding the next point of interest to the optimal path set and marking it as visited; Step S2034: repeat the above step S2033 until all POI data have been visited; Step S2035: Obtain the POI data sequence according to the optimal path set, record the names of each POI data in sequence, and construct a POI corpus.

7. The method for intelligent comprehensive selection of buildings integrating multiple features according to claim 4 is characterized in that: The step S204 specifically includes: The POI corpus is used to train on a Word2Vec model based on CBOW to obtain vectorized representations of POI data of various categories, and the textual POI data is converted into digital data recognizable by a computer to obtain functional semantic features.

8. The intelligent comprehensive selection method of buildings integrating multiple features according to claim 1 is characterized in that: The step S3 specifically includes: Step S301: performing distance weighting on the Delaunay triangulation based on the spatial distance between buildings to obtain a distance weight matrix; Step S302: obtaining a function similarity weight matrix based on the vector cosine similarity relationship between the function semantic features; Step S303: performing weighted processing on the initial graph structure based on the distance weight matrix and the functional similarity weight matrix to obtain a final graph structure.

9. The intelligent comprehensive selection method of buildings integrating multiple features according to claim 1 is characterized in that: The coarse clustering in step S4 specifically includes: Step S401: Initialize K cluster centers based on the coordinates of the geometric center point of the building; Step S402: iterative calculation, assigning each building to the cluster center closest to it, and updating the location of the cluster center; Step S403: using the improved elbow method to determine the K value, by plotting the clustering error sum of squares curves corresponding to different K values, observing the inflection point, and selecting the K value that can retain the regional distribution characteristics to the maximum extent, thereby optimizing the clustering results; Step S404: According to the K value, the original buildings are divided into K categories according to their spatial distribution.

10. The intelligent comprehensive selection method of buildings integrating multiple features according to claim 1 is characterized in that: The step S5 further comprises: During the training process, edges in the graph are randomly deleted to force the model to learn more robust feature representations, thereby improving the generalization ability of the model.

Images