Urban village identification method based on graph neural network and multi-source geographic big data

Abstract

The invention discloses an urban village identification method based on a graph neural network and multi-source geographic big data, and relates to the technical field of geographic images, and the method comprises the steps: obtaining remote sensing data, social perception data and building footprint data of an urban village; based on road network trajectory data, cutting built-up area boundary data through a hierarchical buffer area strategy to obtain a basic evaluation unit; the streetscape image data, the satellite image data, the interest point data, the building attribute data and the building form data are preprocessed respectively, and visual features, social economic features and building form features are constructed; constructing a graph network structure of the graph neural network model; and according to the instance sample group, in combination with a graph network structure of the graph neural network model, constructing and training a GraphSAGE model so as to improve the recognition probability of the villages in the city. According to the method, the spatial interaction and regional relation between the network spatial structure and the multi-dimensional attributes are effectively understood through the GNN.

Description

Technical Field

[0001] This invention relates to the field of geographic image technology, and more specifically, to a method for identifying urban villages based on graph neural networks and multi-source geographic big data. Background Technology

[0002] Urban villages, possessing characteristics of collective land ownership and urban built environment, have led to social problems such as fragmented land use, deterioration of living environment, and safety hazards. In response, urban renewal has been established as a key focus for future work, with the redevelopment of urban villages in megacities and super-large cities serving as pilot projects to improve people's quality of life and promote sustainable urban development.

[0003] Identifying urban villages is a prerequisite for their renovation and transformation. Existing research mostly utilizes high-resolution remote sensing imagery or street view images to extract unique spatial, landscape, and environmental features of urban villages for large-scale identification. In remote sensing imagery, urban villages exhibit significant planar features in terms of spectrum, texture, and morphology; in street view imagery, they show obvious heterogeneity in facade features such as street space, building density, and aging facilities. However, the visual recognition patterns of remote sensing and street view imagery data may lead to a biased identification of urban villages. For example, some old residential areas and old factories have extremely similar appearance features. Furthermore, the inherent socio-economic attributes of urban villages are difficult to extract directly from visual features. For instance, POI data effectively reflects urban functional characteristics; the number of service-related POIs in urban villages is far higher than in ordinary residential areas. The age of buildings and housing prices in a region can significantly reflect the evolution of urban spatial structure; housing prices in urban village areas are significantly lower, and buildings are older. Furthermore, residents' self-construction activities, such as expansions and additions, have resulted in highly identifiable architectural structures and spatial morphological features in urban villages, including building density, height distribution, and spatial pattern. Therefore, integrating multi-source geographic big data such as satellite remote sensing, socio-economic data, and architectural morphology to construct a multi-dimensional urban village identification framework can effectively improve the accuracy of identification based on single visual features.

[0004] Existing methods for identifying urban villages largely rely on satellite imagery and street view images, with limited integration of multimodal data such as regional economic attributes and building morphological features. Furthermore, the understanding of regional spatial relationships and data attribute correlations needs improvement. Therefore, this invention, using a typical city as an example, integrates multi-source geographic big data, including satellite imagery, street view images, points of interest, and vector buildings, to construct an urban village identification framework and interpretability evaluation system based on graph neural networks. This aims to provide technical support for urban renewal and transformation, urban planning and management, and sustainable development.

[0005] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0006] To address the problems in related technologies, this invention proposes a method for identifying urban villages based on graph neural networks and multi-source geographic big data, in order to overcome the aforementioned technical problems existing in existing related technologies.

[0007] Therefore, the specific technical solution adopted by the present invention is as follows: This invention provides a method for identifying urban villages based on graph neural networks and multi-source geographic big data, including: The system acquires remote sensing data, social perception data, and building footprint data for urban villages. The remote sensing data includes satellite imagery data and built-up area boundary data. The social perception data includes road network trajectory data, street view image data, and point of interest data, used to reflect the urban transportation network at different levels. The building footprint data includes building morphology data and building attribute data, used to reflect the geometric attribute information of buildings. Based on road network trajectory data, basic evaluation units are obtained by cropping the boundary data of built-up areas through a hierarchical buffer strategy. Street view image data, satellite image data, point of interest data, building attribute data, and building form data are preprocessed to construct visual features, socio-economic features, and building form features. Based on the basic evaluation unit, a Delaunay triangulation is constructed, and visual features, socio-economic features and architectural morphological features are respectively injected into the nodes of the Delaunay triangulation to construct the graph network structure of the graph neural network model. Based on the example sample group, and combined with the graph network structure of the graph neural network model, the GraphSAGE model is constructed and trained to improve the identification probability of urban villages. We selected precision, accuracy, recall, and F1 score to evaluate the accuracy of the GraphSAGE model in parallel.

[0008] Furthermore, the acquisition of remote sensing data, social perception data, and building footprint data of urban villages; the remote sensing data includes satellite imagery data and built-up area boundary data; the social perception data includes road network trajectory data, street view image data, and point of interest data, used to reflect the urban transportation network at different levels; the building footprint data includes building morphology data and building attribute data, used to reflect the geometric attribute information of buildings, including: Satellite imagery data was obtained using web crawling methods; built-up area boundary data was obtained using existing research methods and open-source data. Road network trajectory data is obtained using the OSM database; street view image data is obtained using the street view database; geographic entity information is recorded using map software APIs to obtain point of interest data; Building morphology data was obtained using the CMAB dataset; building attribute data was obtained using web crawlers on a real estate information platform.

[0009] Furthermore, the basic evaluation units obtained by cropping the built-up area boundary data based on road network trajectory data using a hierarchical buffer strategy include: Based on the built-up area boundary data, the urban village area that conforms to the hierarchical buffer strategy is cropped using cropping software; Based on the urban village area that conforms to the hierarchical buffer strategy, basic evaluation units are generated by combining the centerlines of main roads, secondary roads and branch roads in the road network trajectory data.

[0010] Furthermore, the preprocessing of street view image data, satellite image data, point of interest data, building attribute data, and building morphology data to construct visual features, socioeconomic features, and building morphology features includes: A spatial mapping relationship is constructed, street view image data and satellite image data are preprocessed, street view image features and image features are extracted, and the street view image features and image features are spatially matched to construct visual features. By using a point density estimation algorithm, the density of point of interest data is quantified and reclassified, various density features are extracted, and the housing price and age features in building attribute data are obtained. Combined with point of interest data, socioeconomic features are constructed. By revealing the urban spatial pattern through land development intensity and space utilization efficiency, urban morphology indicators are formed, and building morphological characteristics are constructed by combining building attribute data.

[0011] Furthermore, the construction of the spatial mapping relationship involves preprocessing street view image data and satellite image data, extracting street view image features and image features, and spatially matching the street view image features and image features to construct visual features, including: Geometric correction and image fusion processing are performed on satellite image data, and corresponding image data are cropped according to basic evaluation units. Spectral features, texture features and structural features are extracted from the image data to construct image features. Using the ResNet101 model as the architecture, a semantic segmentation model is constructed. By utilizing the attention mechanism and a local cross-channel interaction strategy, each channel in the semantic segmentation model is adaptively weighted by a one-dimensional convolutional kernel. In addition, by combining open-source samples with self-annotated samples, high-precision samples are provided for the semantic segmentation model, resulting in an improved semantic segmentation model to enhance the extraction accuracy of buildings in urban villages. An improved semantic segmentation model is used to perform detailed qualitative classification of environmental elements in street view image data, dividing urban pixels into urban villages and non-urban villages, and obtaining street view image features. Based on the vehicle driving direction in non-urban villages, a spatial mapping relationship is established between the right-side angle view of the sampling point and the corresponding planar unit to solve the spatial matching problem between point sampling and planar unit. Based on spatial mapping relationships, street view image features and image features are spatially matched to construct visual features.

[0012] Furthermore, the construction of a Delaunay triangulation based on basic evaluation units, and the injection of visual features, socioeconomic features, and architectural morphological features into the nodes of the Delaunay triangulation to construct the graph network structure of the graph neural network model, includes: Using neighborhoods as the basic evaluation unit, a Delaunay triangulation is constructed, and the edges of the Delaunay triangulation are used as the edges of the graph neural network model. Visual features, socioeconomic features, and architectural features are injected into the nodes of the Delaunay triangulation to construct the graph network structure of the graph neural network model.

[0013] Furthermore, the graph network structure for constructing the graph neural network model by injecting visual features, socioeconomic features, and architectural morphological features into the nodes of the Delaunay triangulation includes: Visual features, socioeconomic features, and architectural morphology features are combined using a feature concatenation method to construct a multi-source heterogeneous feature representation; Multi-source heterogeneous feature representations are embedded into the edges of the Delaunay triangulation, and an edge set is generated based on the spatial topological adjacency criterion. By using clipping software to triangulate the edge set and establishing undirected connections between corresponding nodes, a graph network structure for the graph neural network model is formed.

[0014] Furthermore, the step of constructing and training the GraphSAGE model based on the instance sample group and the graph network structure of the graph neural network model to improve the identification probability of urban villages includes: The instance sample groups are labeled and divided into non-urban village and urban village categories, and then labeled using multi-source heterogeneous feature representations to obtain labeled sample groups. Based on labeled sample groups, a three-layer architecture is adopted. Following the principle of progressive aggregation of spatial information, the spatial feature learning from local to macro regions is achieved by expanding the receptive field layer by layer, and the GraphSAGE model is constructed and trained.

[0015] Furthermore, the construction and training of the GraphSAGE model, based on labeled sample groups and employing a three-layer architecture, follows the principle of progressive aggregation of spatial information. It expands the receptive field layer by layer to achieve spatial feature learning from local to macroscopic regions. This includes: The first layer maps the original node features to the intermediate hidden representation and uses the mean aggregator to aggregate the features of neighboring nodes to capture the spatial patterns of the first-order neighborhood, thereby extracting the local spatial relationship between the urban village and the surrounding area. The second layer uses a mean aggregator to aggregate the spatial patterns of the first-order neighborhood into a higher-level representation. By expanding the receptive field of the nodes, the GraphSAGE model captures spatial structural features of a wider area and identifies the clustering distribution and spatial sprawl effect of urban villages. The third layer combines the extracted original node features and local spatial relationships using the softmax function, transforming them into category probabilities and outputting the identification probability of urban villages.

[0016] Furthermore, the calculation formula for the aggregation by the mean aggregator is as follows: ; In the formula, u Represents a node v The neighboring nodes, Represents the first-level node u The hidden representation, N ( v ) represents a node v The neighborhood group, AGG 1 represents an aggregate function. X u Indicates the initial characteristics.

[0017] The beneficial effects of this invention are as follows: 1) This invention utilizes Generative Neural Networks (GNNs) to more effectively capture complex topological relationships, cross-modal data associations, and nonlinear coupling relationships, enabling a better understanding of network spatial structures and spatial interactions and regional connections between multidimensional attributes. Specifically, the GraphSAGE method, through its inductive learning framework and neighborhood aggregation mechanism, effectively handles large-scale heterogeneous urban data and fully leverages limited labeled samples to capture contextual information of urban villages through semi-supervised learning, making it particularly suitable for modeling complex urban systems. Furthermore, the introduction of the GNN-compatible interpretability method GNNExplainer quantifies the relative importance of node connectivity and node features, effectively verifying the intrinsic mechanism of multi-source data synergy.

[0018] 2) This invention uses comprehensive evaluation indicators to analyze the performance of the urban village identification model. Four indicators were selected: accuracy, precision, recall, and F1 score. These indicators quantify the model's overall discrimination accuracy for positive and negative samples, the level of misjudgment control for positive sample prediction results, the ability to avoid missing detections of real positive samples, and the overall balance between misjudgment and missed detection. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of a method for identifying urban villages based on graph neural networks and multi-source geographic big data according to an embodiment of the present invention; Figure 2 This is a diagram of a graph neural network-based framework for identifying urban villages according to an embodiment of the present invention. Figure 3 This is a schematic diagram of visual feature extraction based on remote sensing images and street view images according to an embodiment of the present invention; Figure 4 This is a schematic diagram of parameters for extracting building morphological features according to an embodiment of the present invention; Figure 5 This is an accuracy curve during the training process of the urban village identification model according to an embodiment of the present invention. Figure 6 This is a statistical diagram showing the spatial distribution pattern and area of ​​urban villages in a certain city according to an embodiment of the present invention. Detailed Implementation

[0021] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0022] According to an embodiment of the present invention, a method for identifying urban villages based on graph neural networks and multi-source geographic big data is provided.

[0023] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, a method for identifying urban villages based on graph neural networks and multi-source geographic big data according to an embodiment of the present invention includes: Step S1: Acquire remote sensing data, social perception data, and building footprint data of urban villages; the remote sensing data includes satellite image data and built-up area boundary data; the social perception data includes road network trajectory data, street view image data, and point of interest data, used to reflect the urban transportation network at different levels; the building footprint data includes building morphology data and building attribute data, used to reflect the geometric attribute information of buildings; In this optional embodiment, the acquisition of remote sensing data, social perception data, and building footprint data of urban villages; the remote sensing data includes satellite imagery data and built-up area boundary data; the social perception data includes road network trajectory data, street view image data, and point of interest data, used to reflect the urban transportation network at different levels; the building footprint data includes building morphology data and building attribute data, used to reflect the geometric attribute information of buildings, including: Satellite imagery data was obtained using web crawling methods; built-up area boundary data was obtained using existing research methods and open-source data. Road network trajectory data is obtained using the OSM database; street view image data is obtained using the street view database; geographic entity information is recorded using map software APIs to obtain point of interest (POI) data. Building morphology data was obtained using the CMAB dataset; building attribute data was obtained using web crawlers on a real estate information platform.

[0024] Specifically, three main categories of data—remote sensing, social perception, and building footprint—are used to identify urban villages.

[0025] (1) Remote sensing data includes satellite imagery data and built-up area boundary data.

[0026] The satellite imagery data used was high-resolution remote sensing imagery from Google Earth level-17 in a certain year, with a spatial resolution of approximately 2 meters and a spatial range of 148.5 × 56.1 km, acquired through web crawling. Built-up area boundary data was obtained using existing research methods and open-source data, describing the extent of urban built-up areas nationwide in a certain year, used to determine the approximate boundaries of urban villages.

[0027] (2) Social perception data includes road network trajectory data, street view images, and points of interest data.

[0028] Road network trajectory data was sourced from the OpenStreetMap (OSM) database, which is characterized by its openness, flexibility, and diversity, effectively reflecting the complex traffic networks at different levels within a city. Street view data was obtained from Baidu Maps' street view database, based on OSM's main and secondary roads. The panoramic static image API was used, with a sampling interval of 50m, to acquire street view information from four fields of view (0°, 90°, 180°, and 270°). The study acquired approximately 310,000 street view images of a typical city in a given year, with a resolution of 1024. 700. The POI data comes from the Gaode Map API and records information such as the name, address, coordinates, and category of geographic entities. It is mainly divided into 14 categories, and a total of about 350,000 POI records from typical cities in 2020 were obtained.

[0029] (3) Building footprint data includes building form data and building attribute data.

[0030] The building morphology data comes from the CMAB dataset, which includes geometric attribute information such as roof, height, structure, and outline of all buildings covering the entire area of ​​a typical city in a certain year.

[0031] The building attribute data is based on the Anjuke real estate information platform. It uses web crawling to obtain more than 18,000 housing information entries in typical cities, including latitude and longitude, completion time, building type, and housing price.

[0032] Step S2: Based on road network trajectory data, the boundary data of the built-up area is cropped using a hierarchical buffer strategy to obtain basic evaluation units; In this optional embodiment, the basic evaluation unit obtained by cropping the built-up area boundary data based on road network trajectory data using a hierarchical buffer strategy includes: Based on the built-up area boundary data, the urban village areas that conform to the hierarchical buffer strategy are clipped using clipping software (such as the CLIP tool in ArcGIS software). Based on the urban village area that conforms to the hierarchical buffer strategy, basic evaluation units are generated by combining the centerlines of main roads, secondary roads and branch roads in the road network trajectory data.

[0033] Specifically, the overall framework is as follows: Figure 2 As shown, it is mainly divided into three parts: basic unit division, multi-source feature extraction, and graph neural network recognition. The specific methods are as follows.

[0034] Based on the built-up area boundary data, the CLIP tool in ArcGIS software was used to crop out the approximate areas that correspond to the existence of urban villages. The centerlines of the main, secondary, and branch roads of OSM were used to generate basic evaluation units.

[0035] Specifically, a block is a polygon surrounded by a road network, and therefore can be delineated using buffered road centerlines. Considering that roads of different types and locations have varying widths, the average width of roads at different levels was determined by directly measuring their widths using ArcGIS software with high-resolution Google Earth imagery, based on construction standards and actual conditions. Finally, road space was established based on road buffer zones, and then removed from the study area to generate independent plot units delineated by the road network as the basic units of the study.

[0036] Step S3: Preprocess the street view image data, satellite image data, point of interest data, building attribute data, and building form data respectively to construct visual features, socio-economic features, and building form features; In this optional embodiment, the preprocessing of street view image data and satellite image data, point of interest data and building attribute data and building morphology data to construct visual features, socio-economic features and building morphology features includes: A spatial mapping relationship is constructed, street view image data and satellite image data are preprocessed, street view image features and image features are extracted, and the street view image features and image features are spatially matched to construct visual features. By using a point density estimation algorithm, the density of point of interest data is quantified and reclassified, various density features are extracted, and the housing price and age features in building attribute data are obtained. Combined with point of interest data, socioeconomic features are constructed. By revealing the urban spatial pattern through land development intensity and space utilization efficiency, urban morphology indicators are formed, and building morphological characteristics are constructed by combining building attribute data.

[0037] In this optional embodiment, the construction of spatial mapping relationships, including preprocessing street view image data and satellite image data, extracting street view image features and image features, and spatially matching the street view image features and image features to construct visual features, includes: Geometric correction and image fusion processing are performed on satellite image data, and corresponding image data are cropped according to basic evaluation units. Spectral features, texture features and structural features are extracted from the image data to construct image features. Using the ResNet101 model as the architecture, a semantic segmentation model is constructed. By utilizing the attention mechanism and a local cross-channel interaction strategy, each channel in the semantic segmentation model is adaptively weighted by a one-dimensional convolutional kernel. In addition, by combining open-source samples with self-annotated samples, high-precision samples are provided for the semantic segmentation model, resulting in an improved semantic segmentation model to enhance the extraction accuracy of buildings in urban villages. An improved semantic segmentation model is used to perform detailed qualitative classification of environmental elements in street view image data, dividing urban pixels into urban villages and non-urban villages, and obtaining street view image features. Based on the vehicle driving direction in non-urban villages, a spatial mapping relationship is established between the right-side angle view of the sampling point and the corresponding planar unit to solve the spatial matching problem between point sampling and planar unit. Based on spatial mapping relationships, street view image features and image features are spatially matched to construct visual features.

[0038] Specifically, three types of features were constructed: visual features, socio-economic features, and architectural morphology features, to describe the multidimensional attribute features of basic units.

[0039] (1) Visual Features: The remote sensing images underwent geometric correction and image fusion preprocessing, and the remote sensing images corresponding to the basic evaluation units were cropped out. In the planar feature extraction stage of remote sensing, we selected three major categories of features: spectral, texture, and structure. Among them, spectral features were represented by two indicators: mean and standard deviation; texture features included five indicators: contrast, correlation, energy, entropy, and homogeneity; and structural features were represented by one typical indicator: scale-invariant features (SIFT). In the semantic segmentation stage of street scene images, the DeepLabv3+ model was used to perform detailed qualitative classification of the environmental elements of the street scene, dividing urban pixels into two categories: urban villages and non-urban villages. The DeepLabv3+ model was implemented using the open-source DeepLabV3Plus-ResNet101 model in Python, which was trained using open-source samples. To address the issue of insufficient capture of boundary details of urban village buildings in street view imagery, this invention introduces an attention mechanism. This mechanism employs a local cross-channel interaction strategy, using a one-dimensional convolutional kernel for adaptive weighting of each channel. By combining open-source samples with self-annotated samples, high-precision samples are provided to the DeepLabv3+ semantic segmentation model, thereby improving the extraction accuracy of urban village buildings. Figure 3 As shown in Figure A, the semantic segmentation results of the street view image include urban village buildings, sky, roads, cars, etc. Figure 3 As shown in Figure B, remote sensing image feature extraction includes features such as spectral mean, texture contrast, texture correlation, and scale-invariant SIFT structure. The "urban village" category includes urban village buildings, while the "non-urban village" category includes sky, cars, vegetation, pedestrians, roads, and ordinary buildings. We calculated the pixel proportions of urban villages, sky, vegetation, vertical occlusions such as walls and pillars, and road signs in the street view images, determining five facade feature indices: target ratio, openness, greenness, enclosure, and pictoriality. Simultaneously, due to slight positional deviations between street view sampling points and basic units, it is necessary to determine the spatial relationship between sampling points and basic units. Based on vehicle travel direction, this study establishes a spatial mapping relationship between the 90° right-side view of the sampling point and its corresponding basic unit, thus solving the spatial matching problem between point sampling and area units.

[0040] (2) Socioeconomic characteristics: In the POI feature extraction stage, we used the point density estimation method to quantify the density of 14 POI types (i.e., count the number of POIs in each basic unit and divide by the area of ​​the basic unit to obtain the point density), and classified them into 6 major categories: ① Lifestyle category, including catering, shopping, car services, life services, transportation facilities, etc.; ② Leisure category, including hotels, sports, tourism, etc.; ③ Commerce category, including commerce, finance, etc.; ④ Office category, including companies, government, etc.; ⑤ Medical category, including medical and health care, etc.; ⑥ Culture category, including science and technology culture, etc. At the same time, we constructed information such as house price and house age to construct house attribute characteristics.

[0041] (3) Architectural morphological characteristics: Urban morphological indicators are important quantitative parameters for analyzing urban spatial structure, and can reveal regional spatial patterns from the perspectives of land development intensity and spatial utilization efficiency (e.g., Figure 4 (As shown). This paper constructs four indicators: ① Building Coverage (BSF), which is the ratio of the building footprint area Ai to the total land area A of the plot, directly reflecting the degree of occupancy of hardened surfaces. ② Building Shape Index (BSI), which focuses on characterizing the compactness and complexity of building plan shapes, and is a dimensionless index established by the building outline perimeter Ei and area Ai. ③ Floor Area Ratio (FAR), as a core indicator of the intensity of construction land development, is the ratio of the sum of the floor areas Fi of all floors on the plot to the building footprint area Ai. ④ Space Crowding (SCD), which is the percentage relationship between the total building volume and the volume of the space it occupies. The volume of the space is the product of the highest building height Hmax in the area and the total land area A of the plot, and can be used to evaluate the utilization efficiency of urban three-dimensional space.

[0042] Step S4: Based on the basic evaluation unit, construct a Delaunay triangulation, and inject visual features, socio-economic features and architectural morphology features into the nodes of the Delaunay triangulation to construct the graph network structure of the graph neural network model. In this optional embodiment, the step of constructing a Delaunay triangulation based on basic evaluation units, and injecting visual features, socioeconomic features, and architectural morphological features into the nodes of the Delaunay triangulation to construct the graph network structure of the graph neural network model includes: Using neighborhoods as the basic evaluation unit, a Delaunay triangulation is constructed, and the edges of the Delaunay triangulation are used as the edges of the graph neural network model (GNN). Visual features, socioeconomic features, and architectural features are injected into the nodes of the Delaunay triangulation to construct the graph network structure of the graph neural network model.

[0043] In this optional embodiment, the step of injecting visual features, socioeconomic features, and architectural morphological features into the nodes of the Delaunay triangulation to construct the graph network structure of the graph neural network model includes: Visual features, socioeconomic features, and architectural morphology features are combined using a feature concatenation method to construct a multi-source heterogeneous feature representation; Multi-source heterogeneous feature representations are embedded into the edges of the Delaunay triangulation, and an edge set is generated based on the spatial topological adjacency criterion. The edge set is triangulated using clipping software (DELAUNAY_TRIANGULATION in ArcGIS software), and undirected connecting edges are established between corresponding nodes to form the graph network structure of the graph neural network model.

[0044] Specifically, the graph structure consists of independent nodes and edges, corresponding to geographic entities and the spatial relationships between entities, respectively. This paper uses the basic unit of road network division as the node set V. Each node contains information such as visual features FV, ​​socio-economic features FS, and architectural morphology features FM. Multi-source heterogeneous feature representations are constructed through feature concatenation and embedded into the graph nodes to form a complete attribute graph. An edge set E is generated based on the spatial topological adjacency criterion. The unit boundaries are triangulated using the Delaunay triangulation algorithm (generated using ArcGIS software's DELAUNAY_TRIANGULATION). That is, an undirected connection is established between corresponding nodes if and only if two basic units share the same triangle edge, thus forming the graph structure G(V, E) of the spatial unit. Furthermore, all features are standardized to eliminate the influence of different dimensions of various features on the recognition results, thereby improving the model convergence speed and balancing feature influence.

[0045] Step S5: Based on the instance sample group and the graph network structure of the graph neural network model, construct and train the GraphSAGE model to improve the identification probability of urban villages. In this optional embodiment, the step of constructing and training the GraphSAGE model based on the instance sample group and the graph network structure of the graph neural network model to improve the identification probability of urban villages includes: The instance sample groups are labeled and divided into non-urban village and urban village categories, and then labeled using multi-source heterogeneous feature representations to obtain labeled sample groups. Based on labeled sample groups, a three-layer architecture is adopted. Following the principle of progressive aggregation of spatial information, the spatial feature learning from local to macro regions is achieved by expanding the receptive field layer by layer, and the GraphSAGE model is constructed and trained.

[0046] In this optional embodiment, the construction and training of the GraphSAGE model based on labeled sample groups, using a three-layer architecture and following the principle of progressive aggregation of spatial information, involves expanding the receptive field layer by layer to achieve spatial feature learning from local to macroscopic regions. The first layer maps the original node features to an intermediate hidden representation and uses a mean aggregator to aggregate the features of neighboring nodes to capture the spatial patterns of the first-order neighborhood, extracting the local spatial relationship between the urban village and the surrounding area. The second layer uses a mean aggregator to aggregate the spatial patterns of the first-order neighborhood into a higher-level representation. By expanding the receptive field of the nodes, the GraphSAGE model captures spatial structural features of a wider area, identifying the clustering distribution and spatial sprawl effect of the urban village. The third layer combines the extracted original node features and local spatial relationships through a softmax function, transforming them into class probabilities and outputting the identification probability of the urban village.

[0047] Specifically, sample labeling and model training dataset partitioning are crucial to the recognition capabilities of the Graph Neural Network (GNN) model. Given the unique and difficult-to-identify nature of urban villages, a large number of reliable instance samples were labeled into 0 (negative samples) and 1 (positive samples), representing non-urban village and urban village categories, respectively. Instance samples primarily originated from relevant public projects, government reports, and news case studies as basic units. Simultaneously, this invention also incorporated multi-source data such as high-resolution images, street view images, POI data, nighttime light remote sensing, and housing prices, labeling these data into 0 (negative samples) and 1 (positive samples), again representing non-urban village and urban village categories. Instance samples were derived from relevant public projects, government reports, news case studies, and manually labeled urban village samples specific to this invention. Furthermore, training, validation, and test sets were generated in a 7:1:2 ratio, resulting in 200 urban village samples and 2000 non-urban village samples for GraphSAGE model training.

[0048] The constructed GraphSAGE model adopts a three-layer architecture. GraphSAGE is a commonly used GNN model, designed following the principle of progressive aggregation of spatial information. It achieves spatial feature representation learning from microscopic local to macroscopic regions by progressively expanding the receptive field layer by layer. It can be implemented using the SAGEConv function called from the torch_geometric.nn library in Python. Specifically: ① The first layer is local feature extraction, which maps the original node (self-node) features X to the intermediate hidden representation using a mean aggregator. AGG 1. By capturing the spatial patterns of the first-order neighborhood through the features of the average neighbor nodes, the local spatial association features between urban villages and surrounding areas can be effectively extracted.

[0049] In this optional embodiment, the calculation formula for the aggregation by the mean aggregator is: ; In the formula, u Represents a node v The neighboring nodes, Represents the first-level node u The hidden representation, N ( v ) represents a node v The neighborhood group, AGG 1 represents an aggregate function. X u Indicates the initial characteristics.

[0050] ②The second layer is the aggregation of regional features, through an aggregator. AGG 2. The first-order features are further aggregated into higher-level representations. By expanding the receptive field of the nodes, the spatial structural features of a wider range are captured, enabling the model to perceive spatial association patterns over a larger range, so as to capture the clustering and distribution characteristics and spatial sprawl effects of urban villages.

[0051] ; In the formula, u Represents a node v The neighboring nodes, Represents the first-level node u The hidden representation, N ( v ) represents a node v The neighborhood group, AGG 2 represents an aggregate function.

[0052] Step S6: Select precision, accuracy, recall, and F1 score to evaluate the accuracy of the GraphSAGE model in parallel.

[0053] Specifically, a comprehensive evaluation index was used to analyze the performance of the urban village identification model. Four indicators were selected: accuracy, precision, recall, and F1 score. These indicators quantify the model's overall accuracy in distinguishing positive and negative samples, the level of misjudgment control in the prediction results of positive samples, the ability to avoid missing detections of real positive samples, and the overall balance between misjudgment and missed detection.

[0054] I. Accuracy analysis of model recognition results; such as Figure 5As shown, the training set accuracy (train_acc), validation set accuracy (val_acc), and test set accuracy (test_acc) change with the number of iterations, illustrating the accuracy changes of the training, validation, and test sets during the training of the GraphSAGE model. With the increase of the number of epochs, rapid accuracy growth is observed in the initial stage, reflecting the model's efficient ability to capture the feature space of urban villages. Subsequently, the accuracy curves enter a stable phase, with training, validation, and test accuracies all converging within 97.2%, indicating that the model has constructed a relatively robust feature recognition framework. The best performance occurs at the 789th epoch, indicating that the model training avoids both early underfitting and the risk of overfitting that may occur over a longer period. In summary, the small fluctuations (±2%) between the curves reveal the model's adaptive learning process for the three types of geospatial features constructed in this paper, reflecting the advantages of the GraphSAGE model in the urban village identification task.

[0055] Integrating multi-source heterogeneous information such as remote sensing imagery, street view images, POI distribution, building attributes, and morphological indicators, a multi-dimensional index covering visual features, socioeconomic attributes, and spatial morphological features was established. A framework for urban village identification and interpretation based on GraphSAGE was constructed, demonstrating significant advantages in spatial representation, data fusion, and model construction. Using relevant training samples and parameters, comparisons were made with GCN (Graph Convolutional Network) and GAT (Graph Attention Network) models. Experimental results (as shown in Table 1) show that GraphSAGE achieves an accuracy of 97.2%, representing an improvement of 4.8% and 11.9% in accuracy compared to GAT and GCN models, respectively. The main reason is that GraphSAGE effectively suppresses noise accumulation and alleviates over-smoothing through inductive neighborhood sampling combined with learnable hierarchical aggregation. In contrast, GCN's fixed-mean aggregation easily dilutes discriminative features in heterogeneous neighborhoods; while GAT uses attention to increase the weights of important adjacencies, it is easily driven by local anomalies in sparse and noisy urban networks, limiting its stability and generalization gain. In addition, GraphSAGE's training time per epoch is 3.7 seconds, which is significantly shorter than both GAT and GCN models. This is mainly because it uses small-batch neighborhood sampling and simple aggregation, which reduces computational complexity and memory usage.

[0056] Table 1. Comparison of Training Accuracy and Time for GraphSAGE, GCN, and GAT Models II. Spatial Distribution Characteristics Analysis of Urban Villages: such as Figure 6As shown in Figure A, the overall spatial distribution pattern of urban villages in a certain city exhibits a clear "ring-shaped structure." Based on the identification results, there were a total of 225 urban villages in the city in 2020, covering a total area of ​​61.4 square kilometers, but their spatial distribution showed significant regional heterogeneity. Urban villages were mainly distributed outside the main urban area, accounting for as much as 70.2% of the total, particularly concentrated in the JN, GC, and LH districts. The main urban area had 29.8% of its urban villages, mainly distributed in the YHT and QH districts. Figure 6 Figure B illustrates the distribution characteristics of urban villages in different areas of the city. There are specific differences in the size of urban villages between the main urban area (including areas XH, QH, JY, GL, QX, YHT, etc.) and non-main urban areas: urban villages in the main urban area are generally smaller and more dispersed, while those in the non-main urban area are relatively larger and more widely distributed, with an average area approximately 22.3% larger than those in the main urban area. Comprehensive analysis of spatial distribution and scale characteristics shows that urban villages in this city exhibit a clear regional differentiation pattern: urban villages within the main urban area exhibit geographical characteristics of "low density, small scale, and discrete distribution," while urban villages outside the main urban area exhibit spatial characteristics of "high density, large scale, and clustered distribution."

[0057] III. Multi-Source Geographic Big Data Ablation Experiment Analysis: By constructing a comparative combination experiment and employing feature ablation experiments, the contribution of different types of data to the identification of urban villages was quantitatively evaluated to verify the effectiveness of multi-source geographic big data in urban village identification. Specifically, seven experimental schemes were constructed based on three types of data sources (visual features VI, socioeconomic features SE, and architectural morphology features MO). As shown in Table 2, in the single data source scheme, MO exhibited higher recognition accuracy, and although it only included four variables, its performance was significantly better than the other two single data sources. In the dual data source combination scheme, the recognition accuracy of each combination scheme was generally higher than that of the single data source scheme, confirming the effectiveness and complementarity of multi-source data fusion. Among them, the "VI+MO" combination achieved the best dual-source combination performance, with an accuracy improvement of 5.8% compared to the single MO data source scheme. This indicates that visual features and morphological features have strong complementarity. Ultimately, the model integrating three data sources, "VI+SE+MO", achieved the best recognition results, further improving the accuracy by 2.4% compared to the best dual-source combination scheme. This result proves the complementary role of various data sources in the identification of urban villages: the visual features of remote sensing and street view data construct the basic discrimination framework, the socio-economic features of POI data and building data are integrated into the functional attribute dimension, and the architectural morphological features of building footprint data provide spatial structure analysis capabilities.

[0058] Table 2 Ablation Experiment Results Based on Different Data Combinations In summary, existing urban village identification methods, utilizing the technical solutions described above, are mostly based on satellite imagery and street view images, with limited integration of multimodal data such as regional economic attributes and building morphological features. Furthermore, improvements are needed in understanding regional spatial relationships and data attribute correlations. Therefore, this embodiment aims to take a typical city as an example, integrating multi-source geographic big data such as satellite imagery, street view images, points of interest, and vector buildings to construct an urban village identification framework and interpretability evaluation system based on graph neural networks. This is intended to provide technical support for urban renewal and transformation, urban planning and management, and sustainable development.

[0059] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying urban villages based on graph neural networks and multi-source geographic big data, characterized in that, include: Acquire remote sensing data, social perception data, and building footprint data of urban villages; The remote sensing data includes satellite imagery data and built-up area boundary data; The social perception data includes: road network trajectory data, street view image data, and point of interest data, used to reflect the urban transportation network at different levels; the building footprint data includes: building morphology data and building attribute data, used to reflect the geometric attribute information of buildings. Based on road network trajectory data, basic evaluation units are obtained by cropping the boundary data of built-up areas through a hierarchical buffer strategy. Street view image data, satellite image data, point of interest data, building attribute data, and building form data are preprocessed to construct visual features, socio-economic features, and building form features. Based on the basic evaluation unit, a Delaunay triangulation is constructed, and visual features, socio-economic features and architectural morphological features are respectively injected into the nodes of the Delaunay triangulation to construct the graph network structure of the graph neural network model. Based on the example sample group, and combined with the graph network structure of the graph neural network model, the GraphSAGE model is constructed and trained to improve the identification probability of urban villages. We selected precision, accuracy, recall, and F1 score to evaluate the accuracy of the GraphSAGE model in parallel.

2. The method for identifying urban villages based on graph neural networks and multi-source geographic big data according to claim 1, characterized in that, The acquisition of remote sensing data, social perception data, and building footprint data of urban villages; the remote sensing data includes satellite imagery data and built-up area boundary data; The social perception data includes: road network trajectory data, street view image data, and point of interest data, used to reflect the urban transportation network at different levels; the building footprint data includes: building morphology data and building attribute data, used to reflect the geometric attribute information of buildings, including: Satellite imagery data was obtained using web crawling methods; built-up area boundary data was obtained using existing research methods and open-source data. Road network trajectory data is obtained using the OSM database; street view image data is obtained using the street view database; geographic entity information is recorded using map software APIs to obtain point of interest data; Building morphology data was obtained through the CMAB dataset; building attribute data was obtained through web crawling on a real estate information platform.

3. The method for identifying urban villages based on graph neural networks and multi-source geographic big data according to claim 2, characterized in that, The basic evaluation units obtained by cropping the built-up area boundary data based on road network trajectory data using a hierarchical buffer strategy include: Based on the built-up area boundary data, the urban village area that conforms to the hierarchical buffer strategy is cropped using cropping software; Based on the urban village area that conforms to the hierarchical buffer strategy, basic evaluation units are generated by combining the centerlines of main roads, secondary roads and branch roads in the road network trajectory data.

4. The method for identifying urban villages based on graph neural networks and multi-source geographic big data according to claim 2, characterized in that, The preprocessing of street view image data, satellite image data, point of interest data, building attribute data, and building morphology data to construct visual features, socio-economic features, and building morphology features includes: A spatial mapping relationship is constructed, street view image data and satellite image data are preprocessed, street view image features and image features are extracted, and the street view image features and image features are spatially matched to construct visual features. By using a point density estimation algorithm, the density of point of interest data is quantified and reclassified, various density features are extracted, and the housing price and age features in building attribute data are obtained. Combined with point of interest data, socioeconomic features are constructed. By revealing the urban spatial pattern through land development intensity and space utilization efficiency, urban morphology indicators are formed, and building morphological characteristics are constructed by combining building attribute data.

5. The method for identifying urban villages based on graph neural networks and multi-source geographic big data according to claim 4, characterized in that, The construction of spatial mapping relationships involves preprocessing street view image data and satellite image data, extracting street view image features and image features, and spatially matching the street view image features and image features to construct visual features, including: Geometric correction and image fusion processing are performed on satellite image data, and corresponding image data are cropped according to basic evaluation units. Spectral features, texture features and structural features are extracted from the image data to construct image features. Using the ResNet101 model as the architecture, a semantic segmentation model is constructed. By utilizing the attention mechanism and a local cross-channel interaction strategy, each channel in the semantic segmentation model is adaptively weighted by a one-dimensional convolutional kernel. In addition, by combining open-source samples with self-annotated samples, high-precision samples are provided for the semantic segmentation model, resulting in an improved semantic segmentation model to enhance the extraction accuracy of buildings in urban villages. An improved semantic segmentation model is used to perform detailed qualitative classification of environmental elements in street view image data, dividing urban pixels into urban villages and non-urban villages, and obtaining street view image features. Based on the vehicle driving direction in non-urban villages, a spatial mapping relationship is established between the right-side angle view of the sampling point and the corresponding planar unit to solve the spatial matching problem between point sampling and planar unit. Based on spatial mapping relationships, street view image features and image features are spatially matched to construct visual features.

6. The method for identifying urban villages based on graph neural networks and multi-source geographic big data according to claim 1, characterized in that, The process of constructing a Delaunay triangulation based on basic evaluation units, and injecting visual features, socioeconomic features, and architectural morphological features into the nodes of the Delaunay triangulation to build the graph network structure of the graph neural network model includes: Using neighborhoods as the basic evaluation unit, a Delaunay triangulation is constructed, and the edges of the Delaunay triangulation are used as the edges of the graph neural network model. Visual features, socioeconomic features, and architectural features are injected into the nodes of the Delaunay triangulation to construct the graph network structure of the graph neural network model.

7. The method for identifying urban villages based on graph neural networks and multi-source geographic big data according to claim 6, characterized in that, The graph network structure for constructing a graph neural network model by injecting visual features, socioeconomic features, and architectural morphological features into the nodes of a Delaunay triangulation includes: Visual features, socioeconomic features, and architectural morphology features are combined using a feature concatenation method to construct a multi-source heterogeneous feature representation; Multi-source heterogeneous feature representations are embedded into the edges of the Delaunay triangulation, and an edge set is generated based on the spatial topological adjacency criterion. By using clipping software to triangulate the edge set and establishing undirected connections between corresponding nodes, a graph network structure for the graph neural network model is formed.

8. The method for identifying urban villages based on graph neural networks and multi-source geographic big data according to claim 7, characterized in that, The step of constructing and training the GraphSAGE model based on the instance sample group and the graph network structure of the graph neural network model to improve the identification probability of urban villages includes: The instance sample groups are labeled and divided into non-urban village and urban village categories, and then labeled using multi-source heterogeneous feature representations to obtain labeled sample groups. Based on labeled sample groups, a three-layer architecture is adopted. Following the principle of progressive aggregation of spatial information, the spatial feature learning from local to macro regions is achieved by expanding the receptive field layer by layer, and the GraphSAGE model is constructed and trained.

9. The method for identifying urban villages based on graph neural networks and multi-source geographic big data according to claim 8, characterized in that, The GraphSAGE model, based on labeled sample groups and employing a three-layer architecture, follows the principle of progressive aggregation of spatial information. It expands the receptive field layer by layer to learn spatial features from local to macroscopic regions, and constructs and trains the model as follows: The first layer maps the original node features to the intermediate hidden representation and uses the mean aggregator to aggregate the features of neighboring nodes to capture the spatial patterns of the first-order neighborhood, thereby extracting the local spatial relationship between the urban village and the surrounding area. The second layer uses a mean aggregator to aggregate the spatial patterns of the first-order neighborhood into a higher-level representation. By expanding the receptive field of the nodes, the GraphSAGE model captures spatial structural features of a wider area and identifies the clustering distribution and spatial sprawl effect of urban villages. The third layer combines the extracted original node features and local spatial relationships using the softmax function, transforming them into category probabilities and outputting the identification probability of urban villages.

10. The method for identifying urban villages based on graph neural networks and multi-source geographic big data according to claim 9, characterized in that, The calculation formula for the aggregation by the mean aggregator is as follows: ； In the formula, u Represents a node v The neighboring nodes, Represents the first-level node u The hidden representation, N ( v ) represents a node v The neighborhood group, AGG 1 represents an aggregate function. X u Indicates the initial characteristics.

Images