Urban functional area recognition method and system based on remote sensing image spatial adjacency relationship

Abstract

The present invention relates to the field of remote sensing image processing, and discloses a method and system for identifying urban functional areas based on spatial adjacency of remote sensing images, the method comprising: S1, collecting remote sensing image data of a study area; constructing a ground feature extraction model to extract and classify ground feature features of remote sensing image data, and obtain a feature map; S2, constructing a graph neural network model, using ground feature objects as graph nodes and spatial adjacency of adjacent ground feature objects as edges to construct graph structure data of a network topology structure; the graph neural network model uses a graph neural network graph attention network to extract feature vectors of graph nodes; S3, outputting functional area identification results according to urban functional area cells. The present invention classifies and identifies ground feature elements and divides urban functional area cells based on remote sensing image data, deeply mines the ground feature elements and spatial adjacency relationships within the urban functional area cells, and realizes high-precision identification and division of urban functional areas.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing image processing, and in particular to a method and system for identifying urban functional areas based on remote sensing image spatial adjacency relations. Background Art

[0002] Traditional methods of obtaining information on urban functional areas include field measurements and surveys, but due to high labor costs and long time consumption, it is difficult to effectively implement large-scale urban functional area distribution surveys. With the rapid development of high-resolution remote sensing satellite technology, the expansion of commercial high-resolution satellite data acquisition channels, and the continuous improvement of remote sensing interpretation technology, using satellite remote sensing data to obtain urban functional area information has become a more efficient method. The deep learning of convolutional neural networks can effectively identify remote sensing images. Traditional convolutional neural networks only focus on the extraction and identification of surface features, ignoring the internal elements and spatial structure information of urban functional areas. This leads to insufficient expression of the topological relationship of deep spatial structures, which seriously restricts the identification of urban functional areas. Therefore, how to extract internal deep features based on remote sensing images and identify urban functional areas is a technical problem that needs to be solved urgently. Summary of the invention

[0003] The purpose of the present invention is to provide a method and system for identifying urban functional areas based on the spatial adjacency relationship of remote sensing images. Through a land feature extraction model, remote sensing image data is subjected to land feature feature extraction and classification identification, and urban functional area cells are divided to obtain a feature map containing N land feature objects and divided into urban functional area cells; a graph neural network model uses land feature objects as graph nodes and the spatial adjacency of adjacent land feature objects as edges to construct graph structure data of a network topological structure, thereby realizing the expression and quantification of land feature objects and spatial topological relationships in urban functional area cells and obtaining feature vectors, obtaining all feature vectors in urban functional area cells according to remote sensing image data of a study area, and realizing accurate identification of categories in urban functional area cells.

[0004] The purpose of the present invention is achieved through the following technical solutions:

[0005] A method for identifying urban functional areas based on spatial adjacency of remote sensing images, the method comprising:

[0006] S1. Collect remote sensing image data of the study area; construct a ground feature extraction model to extract and classify ground feature features from remote sensing image data, and obtain a feature map containing N ground feature objects and divided into urban functional area cells;

[0007] S2. Construct a graph neural network model, using the land feature objects as graph nodes and the spatial adjacency of adjacent land feature objects as edges to construct graph structure data of network topology; the graph neural network model uses the graph neural network graph attention network to extract the feature vectors of the graph nodes;

[0008] S3. Collect the feature vectors of all graph nodes according to the city functional area cells and output the functional area recognition results.

[0009] In order to better realize the present invention, the present invention constructs a remote sensing image sample data set containing land feature category labels and urban functional area category labels, and first inputs the land feature extraction model to perform feature extraction, classification and identification of land feature elements, and model training for urban functional area unit division, so as to obtain a feature map corresponding to the remote sensing image sample data set containing N land feature objects and divided into urban functional area units, wherein the land feature categories include primary categories of green space, open space, buildings and roads, the green space includes secondary categories of woodland and grassland, the open space includes secondary categories of parking lots and squares, and the urban functional area categories include commercial areas, residential areas, industrial areas, and mixed functional areas; then the feature map is input into the graph neural network model for model training for urban functional area category recognition.

[0010] Preferably, the terrain feature extraction model is constructed by fusing a convolutional neural network and a Transformer model. The convolutional neural network uses a CNN algorithm to extract feature encoding of fused location information from remote sensing image data and inputs it into the Transformer model. The Transformer model includes several Transformer layers, and the Transformer model performs feature extraction. The terrain feature extraction model adopts a cascade upsampling operation. The cascade upsampling includes several layers of upsampling modules. Each layer of upsampling modules corresponds to a jump connection and upsamples the feature output of the corresponding layer of the convolutional neural network.

[0011] Preferably, the ground feature extraction model divides the urban functional areas into identification cells based on the urban road network, and the method is as follows:

[0012] Identify and extract roads from remote sensing image data for vectorization processing, extract the center lines of the vectorized roads to construct an urban road network, remove redundant points in the urban road network and trim or extend the road center lines to obtain an urban road network containing several closed grids, and use the closed grid divisions as urban functional area identification cells.

[0013] Preferably, the land feature extraction model uses a polygonal building regularization algorithm to perform building outline regularization processing on land feature objects whose land feature categories are buildings.

[0014] Preferably, the graph node in the graph structure data selects the centroid of the ground feature object segmented by the feature graph in step S1 as the node center.

[0015] Preferably, the spatial relationship between the geographical feature objects in the graph structure data is represented by an adjacency matrix G=(V,E), where V represents a vertex set, which is a set with graph nodes as vertices, and E represents a set with spatial adjacency of adjacent geographical feature objects as edges.

[0016] Preferably, the characteristics of the graph node i in the graph structure data are Input graph neural network model corresponding to the feature vector at layer P Aggregate through the graph attention layer to obtain the new feature vector of graph node i , the weight coefficient between graph node i and its adjacent graph node j The expression is as follows:

[0017] , represents the weight coefficient of the feature transformation of the Pth layer, represents the graph node relevance function, Represents the feature vector corresponding to the graph node j in the P layer in the graph neural network model;

[0018] ,in represents a nonlinear activation function, represents the regularization weight coefficient.

[0019] Preferably, the graph neural network model also introduces a multi-head attention mechanism to construct K groups of independent attention layers to concatenate the output results, and the feature vector of graph node i The expression is as follows:

[0020] ,in Represents the splicing operation from 1 to K, represents the weight coefficient calculated by the k-th group of attention mechanism, represents the kth group of learning parameters.

[0021] A system for identifying urban functional areas based on spatial adjacency of remote sensing images comprises a remote sensing image sample data set, a data acquisition module, a ground feature extraction model and a graph neural network model, wherein the remote sensing image sample data of the remote sensing image sample data set are correspondingly annotated with ground feature category labels and urban functional area category labels; the ground feature extraction model uses the remote sensing image sample data set to perform feature extraction, ground feature classification and recognition, and model training for dividing urban functional area units, and obtains a feature map corresponding to the remote sensing image sample data set containing N ground feature objects and divided into urban functional area units; the graph neural network model performs urban functional area classification on the feature map of the ground feature extraction model Model training for energy zone category recognition; the data acquisition module is used to collect remote sensing image data of the study area and input it into the trained land feature extraction model to perform land feature feature extraction and classification recognition and obtain a feature map containing N land feature objects and divided into urban functional area cells. The trained graph neural network model uses land feature objects as graph nodes and the spatial adjacency of adjacent land feature objects as edges to construct graph structure data of the network topology structure. The graph neural network model uses the graph neural network graph attention network to extract the feature vectors of the graph nodes. The graph neural network model collects the feature vectors of all graph nodes according to the urban functional area cells and outputs the functional area recognition results.

[0022] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0023] (1) The present invention uses a land feature extraction model to extract and classify land feature features from remote sensing image data and divide urban functional area cells, thereby obtaining a feature graph containing N land feature objects and divided urban functional area cells; the graph neural network model uses land feature objects as graph nodes and the spatial adjacency of adjacent land feature objects as edges to construct graph structure data of a network topology structure, thereby expressing and quantifying land feature objects and spatial topological relationships in urban functional area cells and obtaining feature vectors; the graph neural network model is based on the corresponding mapping learning and training of all feature vectors in the remote sensing image sample data set and the urban functional area category labels. The present invention can obtain all feature vectors in the urban functional area cells based on the remote sensing image data of the study area and realize accurate identification of the categories in the urban functional area cells.

[0024] (2) The present invention classifies and identifies land features and divides urban functional area cells based on remote sensing image data, deeply explores the spatial adjacency relationship between land features and land features within the cells of urban functional areas, and realizes high-precision identification and division of urban functional areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a flow chart of the method of the present invention;

[0026] Figure 2 It is a schematic diagram of the principle structure of the Transformer model in the embodiment;

[0027] Figure 3 A schematic diagram of the weight coefficients of graph node i and adjacent graph node j obtained by the graph neural network model in the embodiment;

[0028] Figure 4 This is a schematic diagram of the principle flow of an example study area in the embodiment. DETAILED DESCRIPTION

[0029] The present invention is further described in detail below in conjunction with embodiments:

[0030] Example

[0031] like Figure 1 As shown, a method for identifying urban functional areas based on spatial adjacency of remote sensing images includes:

[0032] S1. Collect remote sensing image data of the study area; construct a land feature extraction model to extract and classify land feature features from remote sensing image data, and obtain a feature map containing N land feature objects and divided into urban functional area cells.

[0033] In some embodiments, Figure 2 As shown, the ground feature extraction model of the present invention is constructed by fusing a convolutional neural network and a Transformer model. The convolutional neural network uses a CNN algorithm to perform feature encoding extraction of fused position information on remote sensing image data (using remote sensing image data as input image, the convolutional neural network uses a CNN algorithm to extract the features and position information of the input image and perform fused feature encoding extraction. The features of the input image extracted by the convolutional neural network include spectral features and geometric features. The spectral features include the blue band mean, the green band mean, the red band mean, the near-infrared band mean, the standard deviation of the blue band spectrum, the green band spectrum, the red band light, and the near-infrared band spectrum, the brightness value, the maximum spectral difference and other spectral features. The geometric Features include area, aspect ratio, compactness, boundary index, shape index, rectangle similarity) and input into the Transformer model, the Transformer model includes several Transformer layers, and the Transformer model performs feature extraction; the ground feature extraction model adopts cascade upsampling operation, the cascade upsampling includes several layers of upsampling modules, each layer of upsampling modules corresponds to jump connection and upsamples the feature output of the corresponding layer of the convolutional neural network; preferably, the number of upsampling module layers of the cascade upsampling is the same as the number of feature extraction layers of the convolutional neural network and forms a corresponding relationship, and the upsampling module of the corresponding layer corresponds to the jump connection feature extraction layer for upsampling operation. The ground feature extraction model can segment and extract ground feature elements of various categories in remote sensing image data.

[0034] In some embodiments, the ground feature extraction model divides the urban functional areas into cells based on the urban road network, and the method is as follows:

[0035] Identify and extract roads from remote sensing image data for vectorization processing, extract the centerline of the vectorized road (the extraction of the road centerline follows the principle of the Douglas-Peucker algorithm) to construct an urban road network, remove redundant points in the urban road network and trim or extend the road centerline to obtain an urban road network containing several closed grids, and use the closed grid division as the identification unit of the urban functional area. After extracting the centerline of the vectorized road to construct the urban road network, polygonal approximation is performed on the identified urban road network: the two farthest points A and B of a road centerline are selected from the urban road network and the line segment AB is constructed as the initial chord, and then the vertical distance d from any point C of the line segment AB (for example, the point farthest from the line segment AB) to the line segment AB is calculated in the road centerline; if d is less than the preset threshold, the line segment AB is regarded as an approximation of the curve; if d is greater than the threshold, the road centerline is divided into two parts, AC and BC, with point C as the segmentation point, and the above method is recursively executed on these two parts until the d of all parts meets the threshold condition. Finally, the polyline formed by connecting all the segmentation points is obtained to obtain a regular linear urban road network. The topological connectivity of the urban road network is further checked, and the data of abnormal road networks in the urban road network is pruned or extended to obtain closed road network data, and the closed grid division is used as the identification cell of the urban functional area.

[0036] In some embodiments, the feature extraction model uses a polygonal building regularization algorithm to regularize the building outline for the feature object whose feature category is a building. The polygonal building regularization algorithm is as follows: First, connect the starting point and the end point of the building outline to form a straight line. Then, calculate the distance from all key points on the building outline to this straight line, and find the maximum distance value among them. ; Then use this maximum distance value With the preset threshold Compare: If the maximum distance value Greater than or equal to threshold , then the maximum distance value The coordinate point of the building is used as the boundary to split the building outline into two parts; if the maximum distance value Less than threshold , then remove the middle point on the building outline; the simplified curve is regarded as the final line segment of the building outline. Repeat the above steps, continuously connect and retain key points until the simplification process of the entire building outline is completed, and finally obtain the building with a simplified polygonal outline.

[0037] Preferably, in the graph structure data of the present invention, the centroid (the centroid of the geography feature object can be replaced by the center point of the geography feature object area) or the center of the geography feature object segmented by the feature graph in the step S1 is used as the node center; if the geography feature object is a building, the centroid (the centroid can be replaced by the center point of the building outline area) or the center after the regularization processing of the building outline is selected.

[0038] S2. Construct a graph neural network model, and construct graph structure data of network topology structure with land feature objects as graph nodes and spatial adjacency of adjacent land feature objects as edges (graph structure data describes the spatial pair relationship of land feature objects as a graph structure by constructing a regional adjacency graph, and completes the spatial structure expression and quantification of land feature objects through graph neural network). The spatial relationship between land feature objects in the graph structure data is represented by the adjacency matrix G=(V,E), where V represents the vertex set, which is a set with graph nodes as vertices, and E represents a set with spatial adjacency of adjacent land feature objects as edges; the adjacency matrix G is an N-order square matrix (N is the total number of vertices, that is, the total number of graph nodes), and the element value of the matrix reflects whether there is an edge between vertices. If vertex i has an edge pointing to vertex j, the corresponding element value in the adjacency matrix is ​​1, otherwise it is 0, thereby representing the adjacency relationship between land feature objects.

[0039] The graph neural network model uses the graph neural network graph attention network to extract the feature vector of the graph node. The features of graph node i in the graph structure data (assuming that the center of the graph node being updated is graph node i) Input graph neural network model corresponding to the feature vector at layer P Aggregate through the graph attention layer to obtain the new feature vector of graph node i ,like Figure 3 As shown, the weight coefficient of graph node i and adjacent graph node j The expression is as follows:

[0040] , Indicates the Pth layer (such as Figure 4 As shown, the graph neural network model includes several layers of graph convolution, and the Pth layer is the weight coefficient of the feature transformation of the Pth layer of the graph neural network model. represents the graph node relevance function, Represents the feature vector corresponding to the graph node j in the P layer in the graph neural network model; this embodiment designs a fully connected layer to output the correlation between the adjacent network node i and the network node j and uses the Softmax function to normalize the correlation calculated for all adjacent network nodes to obtain the new feature vector of the graph node i ;

[0041] ,in represents a nonlinear activation function, represents the regularization weight coefficient.

[0042] In some embodiments, the graph neural network model also introduces a multi-head attention mechanism to construct K groups of independent attention layers to splice the output results. At this time, the feature vector of graph node i is The expression is as follows:

[0043] ,in Represents the splicing operation from 1 to K, represents the weight coefficient calculated by the k-th group of attention mechanism, Represents the kth group of learning parameters. The attention layer of the graph neural network model increases the dimension of the adaptive edge weight coefficient and reduces the learning parameters by introducing the multi-head attention mechanism. The multi-head attention mechanism can distribute attention to multiple related features of the central network node and the neighboring network node, thereby improving the learning ability of the model.

[0044] S3. Collect the feature vectors of all graph nodes according to the urban functional area cells and output the functional area recognition results. The present invention uses the remote sensing image sample data set to successively undergo model training of the ground feature extraction model and the graph neural network model. The graph neural network model uses the remote sensing image sample data set to construct a corresponding mapping relationship between the feature vectors of all graph nodes in the urban functional area cells with remote sensing image samples and the urban functional area label categories (model learning training). The graph neural network model can judge and output the functional area recognition results based on the feature vectors of all graph nodes in the urban functional area cells (the functional area recognition results include commercial areas, residential areas, industrial areas, and mixed functional areas, that is, the output urban functional area cells belong to the functional area categories of commercial areas, residential areas, industrial areas, or mixed functional areas). Figure 4 As shown, this embodiment uses Jilin-1 high-resolution remote sensing image data as the data source, and selects a certain area in Ningbo City as the research area. Figure 4 The construction of the graph structure data (i.e., obtaining the graph structure data in step S2) includes the following steps: inputting remote sensing image data, using the feature extraction model to classify and identify the features (identify and segment the features in the remote sensing image) to obtain a feature map containing N features and divided into urban functional area cells, and then constructing the graph structure data of the network topology structure according to the method in step S2. Figure 4 As shown, the graph neural network model extracts the feature vectors of the graph nodes and outputs them.

[0045] Both the ground feature extraction model and the graph neural network model of the present invention need to be trained, and the method is as follows: a remote sensing image sample data set containing ground feature category labels and urban functional area category labels is constructed, and the data set is first input into the ground feature extraction model for feature extraction, ground feature classification and recognition, and model training for urban functional area unit division, so as to obtain a feature map corresponding to the remote sensing image sample data set containing N ground feature objects and divided into urban functional area units, wherein the ground feature categories include primary categories of green space, open space, buildings and roads, the green space includes secondary categories of woodland and grassland, the open space includes secondary categories of parking lots and squares, and the urban functional area categories include commercial areas, residential areas, industrial areas, and mixed functional areas; then the feature map is input into the graph neural network model for model training for urban functional area category recognition, and the ground feature extraction model of step S1 and the graph neural network model of step S2 of the present invention complete the model training according to the above method and adopt the optimal ground feature extraction model and graph neural network model after training. The remote sensing image sample data in the remote sensing image sample data set are subjected to radiation correction, orthorectification, image fusion, etc., which belong to preprocessing operations. Remote sensing images with high spatial resolution and rich spectral information are selected as remote sensing image sample data. The remote sensing image sample data are annotated with primary categories of green space, open space, buildings and roads. At the same time, secondary categories are annotated under the annotated primary categories. Then the remote sensing image sample data carry category label information, and the annotated remote sensing image sample data are rasterized. The ground feature extraction model of the present invention uses the remote sensing image sample data set model to use a small learning rate in the early stage of model training to promote the stable start of the model, and then switches to a preset learning rate to accelerate and stabilize the convergence process of the model; at the same time, by continuously monitoring the performance of the model on the validation set, including the loss value and accuracy, once the accuracy of the validation set reaches a preset threshold and the reduction in the loss value in several consecutive iterations is lower than the given threshold, the current optimal model state is saved. Based on the weight file of the optimal ground feature extraction model obtained, the sliding window prediction method is used to predict the results. Specifically, the image is cropped with a given overlap and predicted block by block. Each time the prediction result only retains the central area, and the prediction result of the edge of the image is discarded, so as to obtain the prediction result of the large-size remote sensing image without splicing traces.

[0046] A system for identifying urban functional areas based on spatial adjacency of remote sensing images comprises a remote sensing image sample data set, a data acquisition module, a ground feature extraction model and a graph neural network model, wherein the remote sensing image sample data of the remote sensing image sample data set are correspondingly annotated with ground feature category labels and urban functional area category labels; the ground feature extraction model uses the remote sensing image sample data set to perform feature extraction, ground feature classification and recognition, and model training for dividing urban functional area units, and obtains a feature map corresponding to the remote sensing image sample data set containing N ground feature objects and divided into urban functional area units; the graph neural network model performs urban functional area classification on the feature map of the ground feature extraction model Model training for energy zone category recognition; the data acquisition module is used to collect remote sensing image data of the study area and input it into the trained land feature extraction model to perform land feature feature extraction and classification recognition and obtain a feature map containing N land feature objects and divided into urban functional area cells. The trained graph neural network model uses land feature objects as graph nodes and the spatial adjacency of adjacent land feature objects as edges to construct graph structure data of the network topology structure. The graph neural network model uses the graph neural network graph attention network to extract the feature vectors of the graph nodes. The graph neural network model collects the feature vectors of all graph nodes according to the urban functional area cells and outputs the functional area recognition results.

[0047] 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 and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for identifying urban functional areas based on spatial adjacency of remote sensing images, characterized by: The methods include: S1. Collect remote sensing image data of the study area; construct a land feature extraction model to extract and classify land feature features from remote sensing image data, and obtain a feature map containing N land feature objects and divided into urban functional area cells; the land feature extraction model is constructed by fusion of convolutional neural network and Transformer model, the convolutional neural network uses CNN algorithm to extract feature encoding of fused position information from remote sensing image data and inputs it into Transformer model, the Transformer model includes several Transformer layers, and the Transformer model performs feature extraction; the land feature extraction model adopts cascade upsampling operation, the cascade upsampling includes several layers of upsampling modules, each layer of upsampling module corresponds to jump connection and upsamples the feature output of the corresponding layer of convolutional neural network; the land feature extraction model is based on urban road network to divide urban functional area identification cells, the method is as follows: Identify and extract roads from remote sensing image data for vector processing, extract the center lines of the vectorized roads to construct an urban road network, remove redundant points in the urban road network and trim or extend the road center lines to obtain an urban road network containing several closed grids, and use the closed grid division as the identification unit of the urban functional area; S2. Construct a graph neural network model, using the land feature objects as graph nodes and the spatial adjacency of adjacent land feature objects as edges to construct graph structure data of network topology; the graph neural network model uses the graph neural network graph attention network to extract the feature vectors of the graph nodes; S3. Collect the feature vectors of all graph nodes according to the city functional area cells and output the functional area recognition results.

2. The method for identifying urban functional areas based on spatial adjacency of remote sensing images according to claim 1, characterized in that: To construct a remote sensing image sample data set containing land feature category labels and urban functional area category labels, the land feature extraction model is first input for feature extraction, land feature classification and recognition, and model training for urban functional area unit division. The remote sensing image sample data set is obtained to correspond to a feature map containing N land feature objects and divided into urban functional area units. The land feature categories include the primary categories of green space, open space, buildings and roads, the secondary categories of green space include woodland and grassland, the secondary categories of open space include parking lots and squares, and the categories of urban functional areas include commercial areas, residential areas, industrial areas, and mixed functional areas. Then the feature map is input into the graph neural network model for model training for urban functional area category recognition.

3. The method for identifying urban functional areas based on spatial adjacency of remote sensing images according to claim 2, characterized in that: The ground feature extraction model uses a polygonal building regularization algorithm to perform building outline regularization processing on ground feature objects whose ground feature categories are buildings.

4. The method for identifying urban functional areas based on spatial adjacency of remote sensing images according to claim 3 is characterized by: The graph node in the graph structure data selects the centroid of the ground feature object segmented by the feature graph in step S1 as the node center.

5. The method for identifying urban functional areas based on spatial adjacency of remote sensing images according to claim 1, characterized in that: The spatial relationship between the geographical feature objects in the graph structure data is represented by the adjacency matrix G=(V,E), where V represents a vertex set, which is a set with graph nodes as vertices, and E represents a set with the spatial adjacency of adjacent geographical feature objects as edges.

6. The method for identifying urban functional areas based on spatial adjacency of remote sensing images according to claim 1 or 5, characterized in that: Features of graph node i in the graph structure data Input graph neural network model corresponding to the feature vector at layer P Aggregate through the graph attention layer to obtain the new feature vector of graph node i , the weight coefficient between graph node i and adjacent graph node j The expression is as follows: , represents the weight coefficient of the feature transformation of the Pth layer, represents the graph node relevance function, Represents the feature vector corresponding to the graph node j in the P layer in the graph neural network model; ,in represents a nonlinear activation function, represents the regularization weight coefficient.

7. The method for identifying urban functional areas based on spatial adjacency of remote sensing images according to claim 6, characterized in that: The graph neural network model also introduces a multi-head attention mechanism to construct K groups of independent attention layers to splice the output results. The feature vector of graph node i The expression is as follows: ,in Represents the splicing operation from 1 to K, represents the weight coefficient calculated by the k-th group of attention mechanism, represents the kth group of learning parameters.

8. An urban functional area recognition system based on remote sensing image spatial adjacency relationship for implementing the urban functional area recognition method of claim 1, characterized in that: It includes a remote sensing image sample data set, a data acquisition module, a ground feature extraction model and a graph neural network model, wherein the remote sensing image sample data of the remote sensing image sample data set are correspondingly annotated with ground feature category labels and urban functional area category labels; the ground feature extraction model uses the remote sensing image sample data set to perform feature extraction, ground feature classification and recognition, and model training for urban functional area unit division, and obtains a feature map corresponding to the remote sensing image sample data set containing N ground feature objects and divided into urban functional area units; the graph neural network model performs model training for urban functional area category recognition on the feature map of the ground feature extraction model; the data acquisition module is used to collect remote sensing image data of the study area and input it into the trained ground feature extraction model to perform ground feature feature extraction and classification and recognition and obtain a feature map containing N ground feature objects and divided into urban functional area cells; the trained graph neural network model uses ground feature objects as graph nodes and the spatial adjacency of adjacent ground feature objects as edges to construct graph structure data of a network topology structure; the graph neural network model uses the graph neural network graph attention network to extract feature vectors of graph nodes; the graph neural network model collects feature vectors of all graph nodes according to urban functional area cells and outputs functional area recognition results.

Images