Urban land utilization function area identification method based on remote sensing image, terminal and medium

Abstract

The invention provides an urban land utilization function area identification method based on a remote sensing image. The method comprises the following steps: constructing an initial feature set, a spatial proximity graph and a semantic similarity graph corresponding to an image, constructing sample data corresponding to the image in combination with functional region information of the image, training a preset model by using the sample data, and obtaining a trained functional region recognition model; and obtaining a functional area identification result of the image based on the trained functional area identification model. According to the method provided by the invention, the association features among the image spots in the image can be deeply mined, and deep learning is carried out on the topological relation of the image structure based on spatial proximity and semantic similarity by using the image neural network, so that the dependence on multi-source auxiliary data is greatly reduced, the training convenience and efficiency of the functional area recognition model are improved, and the recognition efficiency of the functional area is improved. And engineering application of the technology is effectively promoted.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a method, terminal, and computer storage medium for identifying urban land use functional zones based on remote sensing images. Background Technology

[0002] Different urban land use functional zones are composed of different geographical elements; for example, industrial zones typically include geographical elements such as factories, internal roads, and green spaces; while residential zones are mainly composed of geographical elements such as buildings, community green spaces, internal roads, and swimming pools.

[0003] In the existing functional area identification process, it is usually based on remote sensing image data and combined with multi-source geographic information data (such as POI and traffic data) and a manual interpretation method for comprehensive identification. Although this method is intuitive, it is inefficient and highly subjective. Although there are now automated intelligent identification methods such as object-oriented segmentation and classification methods or convolutional neural networks, existing methods often rely only on the spectral, texture, and shape information within a single tile, while ignoring the overall spatial and semantic information within the image. This results in insufficient perception capability in complex scenes, weak cross-regional transfer capability, and limited engineering accuracy. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide a method, terminal and computer storage medium for identifying urban land use functional zones based on remote sensing images, to solve the problems of existing technologies that only focus on the internal structure of a single image patch without fully exploring the overall spatial and semantic information inside the image, resulting in inaccurate functional zone information, large errors and insufficient transferability of the identification method.

[0005] To achieve the above and other related objectives, the present invention provides a method for identifying urban land use functional zones based on remote sensing imagery in a first aspect, comprising: extracting image feature vectors of various land parcels in a remote sensing imagery according to preset image feature factors to construct an initial feature set corresponding to the imagery; constructing a spatial proximity map and a semantic similarity map corresponding to the imagery by performing spatial proximity analysis and image feature similarity analysis on the land parcels; constructing sample data corresponding to the imagery based on the initial feature set, spatial proximity map, semantic similarity map, and functional zone information; training a preset functional zone identification model based on the sample data, and using the trained functional zone identification model to perform the functional zone identification process of the image to be identified, thereby obtaining the corresponding functional zone identification result; wherein, the functional zone identification model is a hierarchical graph neural network model constructed based on the spatial proximity map and the semantic similarity map.

[0006] The present invention provides another method for identifying urban land use functional zones based on remote sensing imagery in a second aspect, comprising: acquiring remote sensing imagery covering the functional zones, wherein the remote sensing imagery contains geographic barrier elements; extracting image feature vectors corresponding to each patch in the imagery according to preset image feature factors to construct an initial feature set corresponding to the imagery; performing spatial proximity analysis, image feature similarity analysis, and geographic barrier analysis on each patch in the imagery to construct spatial proximity map, semantic similarity map, and barrier map structure corresponding to the imagery; constructing sample data based on the spatial proximity map, semantic similarity map, barrier map structure, and functional zone information of each imagery; training a preset functional zone identification model based on the sample data; acquiring an image to be identified and extracting geographic barrier elements from the imagery; and identifying the functional zones using the trained functional zone identification model based on the image to be identified and the geographic barrier elements to obtain the identification results of the functional zones; wherein the functional zone identification model is a hierarchical graph neural network model constructed based on the spatial proximity map, semantic similarity map, and barrier map structure.

[0007] In one embodiment of the second aspect of this application, the method for obtaining the land parcel includes: performing image segmentation on the remote sensing image in an object-oriented manner to obtain each parcel; and removing parcels that overlap with geographic barrier elements from the segmented parcels, and using the remaining parcels as the land parcel.

[0008] In one embodiment of the second aspect of this application, the method for obtaining the spatial proximity map includes: extracting the center point of each land parcel and setting it as a node in the spatial proximity map; obtaining the spatial neighbor nodes corresponding to the nodes by calculating the degree of adjacency between the spatial positions of each node; and constructing the edge between the node and the corresponding spatial neighbor node and setting it as an edge in the spatial proximity map.

[0009] In one embodiment of the second aspect of this application, the method of obtaining the semantic similarity graph includes: setting the center point of each land parcel as a node in the semantic similarity graph; setting the image feature vector of the land parcel as the image feature vector of the node corresponding to the land parcel; obtaining the semantic neighbor node corresponding to the node by calculating the similarity between the image feature vectors of each node; and constructing the edge between the node and the corresponding semantic neighbor node as the edge in the semantic similarity graph.

[0010] In one embodiment of the second aspect of this application, the method of constructing the barrier graph structure includes: setting the center point of each land parcel as a node in the barrier graph structure; obtaining the barrier neighbor node corresponding to the node by determining the positional relationship between any two nodes and the geographical barrier element based on the spatial position of the node and the spatial position of the geographical barrier element; and constructing the edge between the node and the corresponding barrier neighbor node as the edge in the barrier graph structure.

[0011] In one embodiment of the second aspect of this application, the functional area recognition model includes: an input layer for inputting the initial feature set, the spatial proximity map, the semantic similarity map, and the barrier map structure of the image; a first graph neural network layer connected to the input layer for performing enhancement processing on the initial feature set based on the spatial proximity map and the barrier map structure to obtain a first intermediate feature distribution that integrates spatial context information and geographical barrier information; a second graph neural network layer connected to the input layer and the first graph neural network layer for performing feature enhancement processing on the first intermediate feature distribution based on the semantic similarity map and the barrier map structure to obtain a second intermediate feature distribution that incorporates global semantic context information and geographical barrier information; a feature fusion layer connected to the first graph neural network layer and the second graph neural network layer for fusing the features of the first intermediate feature distribution and the second intermediate feature distribution to obtain a fused feature distribution; and an output layer connected to the feature fusion layer for performing classification prediction on various map patches in the sample data based on the fused feature distribution to obtain the functional area recognition prediction result.

[0012] In one embodiment of the second aspect of this application, the first graph neural network layer employs a spatial graph convolutional layer, and the second graph neural network layer employs a semantic graph convolutional layer; based on the spatial graph convolutional layer, the following steps are performed: extracting the spatial adjacency matrix corresponding to the spatial proximity graph, and extracting the barrier adjacency matrix of the barrier graph structure; performing a de-barriering process on the spatial proximity matrix based on the barrier adjacency matrix to obtain a new spatial proximity matrix; adding a self-connection matrix to the new spatial proximity matrix and calculating the corresponding degree matrix; performing a symmetric normalized graph convolution on the initial feature set based on the degree matrix to obtain a first intermediate feature distribution; and, based on the semantic graph convolutional layer, the following steps are performed: extracting the semantic adjacency matrix corresponding to the semantic similarity graph; performing a de-barriering process on the semantic adjacency matrix based on the barrier adjacency matrix to obtain a new semantic adjacency matrix; adding a self-connection matrix to the new semantic adjacency matrix and calculating the corresponding degree matrix; performing a symmetric normalized graph convolution on the first intermediate feature distribution based on the degree matrix to obtain a second intermediate feature distribution.

[0013] In one embodiment of the second aspect of this application, the first graph neural network layer employs a spatial graph attention layer, and the second graph neural network layer employs a semantic graph attention layer; based on the spatial graph attention layer, the following steps are performed: obtaining the initial feature distribution of the initial feature set after linear transformation under different attention mechanisms; constructing a spatial attention coefficient between a node and its corresponding spatial neighbor nodes in the initial feature distribution after linear transformation, combined with the node's corresponding barrier information, to enhance attention on the node's initial feature vector after linear transformation based on the spatial attention coefficient, thereby obtaining a first intermediate feature distribution under comprehensive barrier information; and, based on the semantic graph attention layer, the following steps are performed: obtaining the first intermediate feature distribution after linear transformation under different attention mechanisms; constructing a semantic attention coefficient between a node and its corresponding semantic neighbor nodes in the first intermediate feature distribution after linear transformation, combined with the node's corresponding barrier information, to enhance attention on the node's first intermediate feature vector after linear transformation based on the semantic attention coefficient, thereby obtaining a second intermediate feature distribution under comprehensive barrier information.

[0014] In one embodiment of the second aspect of this application, the functional area identification process of the image to be identified includes: extracting image feature vectors of various land parcels in the image to be identified according to preset image feature factors to obtain an initial feature set corresponding to the image to be identified; constructing a spatial proximity map and a semantic similarity map corresponding to the image to be identified by performing spatial proximity analysis and image feature similarity analysis on each land parcel; and identifying the image to be identified using the trained functional area identification model based on the initial feature set, the spatial proximity map, the semantic similarity map, and the barrier map structure of the image to be identified.

[0015] In a third aspect, this application provides a terminal comprising: a processor and a memory, wherein the memory and the processor are communicatively connected; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the terminal performs the urban land use functional zone identification method based on remote sensing imagery as described above.

[0016] Finally, the present invention also provides a computer storage medium storing a computer program that, when executed by a processor, implements the urban land use functional zone identification method based on remote sensing imagery as described above.

[0017] As described above, the urban land use functional zone identification method, terminal, and computer storage medium based on remote sensing imagery provided by this invention deeply mines the correlation features between various image patches within the functional zone, and utilizes graph neural networks to perform deep learning on the graph structure topological relationships based on spatial proximity and semantic similarity. This not only reduces the dependence on multi-source auxiliary data such as POIs and street views, but also greatly reduces the amount of sample data required for model training. That is, end-to-end deep learning model training can be achieved based solely on remote sensing imagery and a small amount of sample data, resulting in a functional zone identification model with good recognition performance. In addition, the method of this application greatly reduces the tedious manual parameter tuning in traditional model training methods, effectively promoting the engineering application of the technology. Attached Figure Description

[0018] Figure 1 The diagram shows the distribution of functional areas in one embodiment of this application.

[0019] Figure 2 This is shown as an embodiment of the urban land use functional zone identification based on remote sensing imagery described in this application.

[0020] A flowchart illustrating the method;

[0021] Figure 3 This is a schematic diagram showing the same ground features being blocked by geographical elements in one embodiment of this application;

[0022] Figure 4 This is shown as an urban land use functional zone identification based on remote sensing imagery in another embodiment of this application.

[0023] A flowchart illustrating another method;

[0024] Figure 5 The diagram shown is a structural schematic of the terminal described in one embodiment of this application. Detailed Implementation

[0025] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0026] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention, and are therefore not intended to be considered as drawings.

[0027] To facilitate understanding of the technical solutions and effects in this application, the following brief description is provided:

[0028] Urban land use functional zones (hereinafter referred to as "functional zones") are regional types classified according to dominant functions in urban planning and management. They mainly include residential areas, commercial areas, industrial areas, cultural and educational areas, and leisure areas. Each type of functional zone is not a single land feature, but rather a combination of a series of spatially and functionally related geographical elements. For example, an industrial zone consists of elements such as factories, internal roads, green spaces, and storage yards; a residential zone consists of elements such as buildings, community green spaces, internal roads, and swimming pools.

[0029] Spatial proximity map: A graph structure used to represent the spatial adjacency relationship between various geographic elements (spatially adjacent) within a functional area; for example, factory buildings in an industrial area are adjacent to internal road patches, while factory buildings in an industrial area are not adjacent to river patches, etc.; nodes in the spatial proximity map represent the center point of the corresponding geographic element, and edges represent the spatial adjacency relationship of geographic elements.

[0030] Semantic similarity graph: A graph structure used to characterize the semantic similarity between geographic elements (spatially not necessarily adjacent) within a functional area in terms of image feature dimensions; where semantic information includes remote sensing image features; for example, for buildings of the same style that are discretely (non-adjacent) in the same commercial area, the texture features of their roofs are often similar; or for residential buildings that are discretely (non-adjacent) in the same residential area, the spectral features of their roofs are often similar; in the semantic similarity graph, nodes represent the center point of the corresponding geographic element, and edges represent the semantic similarity relationship between geographic elements.

[0031] Barrier structure diagram: This is a graph structure used to characterize whether geographical features are spatially separated by geographical barrier features such as main roads and rivers. For example, for two buildings, if they are both located on the same side of a main river, it indicates that they are not geographically separated; while if they are located on different sides of the main river, it indicates that they are geographically separated. The nodes in the diagram represent the corresponding geographical features, and the edges represent whether the geographical features are isolated by spatial barrier features.

[0032] This invention first provides a method for identifying urban land use functional zones based on remote sensing imagery, which is used to accurately extract functional zone information in complex urban environments; wherein, the functional zone information is information related to the spatial distribution and type of functional zones, including but not limited to distribution area, area size and functional zone type.

[0033] Please see Figure 1 The diagram shows the distribution of functional areas in remote sensing imagery; for example... Figure 1As shown, the polygons in the image represent the boundaries of functional areas, and different colors represent different types of functional areas; specifically, functional areas 1 to 6 in the image correspond to industrial area, industrial area, cultural and educational area, residential area, residential area and leisure area, respectively.

[0034] Please see Figure 2 The diagram illustrates the execution flow of the urban land use functional zone identification method based on remote sensing imagery according to an embodiment of this application; as shown below. Figure 2 As shown, the method includes the following steps:

[0035] S100: Acquire remote sensing images covering the functional area, which contain land parcels corresponding to geographic features;

[0036] The land parcels are the sub-images obtained by segmenting remote sensing images (hereinafter referred to as "images") according to geographic elements, and each land parcel corresponds to a geographic element.

[0037] Specifically, the methods for obtaining map features for each land parcel include:

[0038] Based on the spectral, texture, and morphological information of images, object-oriented segmentation is performed on the images, thereby enabling the rapid and convenient acquisition of land parcels corresponding to geographic features.

[0039] In a more specific embodiment, the method for obtaining the land parcel map includes:

[0040] Superpixel segmentation is used to segment an image, which involves aggregating pixels with similar spectral information and spatial proximity into the same patch, thereby dividing the entire image into different patch blocks, denoted as S = {s1, s2, ..., s}. n}; where S represents the set of land parcels; s i Let represent the i-th land parcel, and n be the total number of land parcels in the image.

[0041] The superpixel segmentation algorithm includes, but is not limited to, the Simple Linear Iterative Clustering (SLIC) algorithm.

[0042] It should be noted that before performing superpixel segmentation, it is usually necessary to set the expected number K of superpixels to determine the superpixel scale; that is, the larger the K value, the smaller the superpixel. In the embodiments of this application, the expected number K is determined based on the spatial resolution of the image and the complexity of the functional areas within the corresponding region of the image. Specifically, the higher the spatial resolution of the image and the more complex the functional areas, the larger the K value, and vice versa. The complexity of the functional areas refers to the number of geographic features contained within the functional areas; that is, the more geographic features, the more complex the functional areas.

[0043] This application divides the image into different land parcels, replacing the original image pixels as the basic units for subsequent processing. This not only greatly reduces the amount of data in subsequent processing and lowers the computational complexity, but also ensures that each basic unit has relatively consistent image features and retains the boundary information of geographic elements, thereby ensuring the accuracy of functional area identification results.

[0044] S200: Based on preset image feature factors, extract the corresponding image feature vectors of each patch to construct the initial feature set corresponding to the image.

[0045] Among them, image feature factors are used to characterize the feature information of land parcels on images, including one or more of spectral feature factors, texture feature factors, and morphological feature factors.

[0046] Specifically, the spectral feature factors include the values ​​or ratios of land parcel patches in characteristic bands such as red, green, blue, and / or near-infrared, used to reflect the spectral absorption characteristics of land features; the texture feature factors include information such as the contrast, correlation, energy, or homogeneity of pixel values ​​within land parcel patches, used to characterize the roughness, regularity, and other texture structure information of land features; the morphological feature factors include the area, perimeter, aspect ratio, or compactness (perimeter) of land parcel patches. 2 The area (e.g., area) is used to distinguish the shapes of different land features, such as linear features (e.g., roads), block features (e.g., buildings), and irregular features (e.g., green spaces).

[0047] Based on the calculation methods corresponding to each image feature factor, feature values ​​of land parcels in the image are extracted respectively; the feature values ​​of land parcels corresponding to different image feature factors are summarized to generate an initial feature vector of land parcels; the initial feature vector is a multi-dimensional feature vector, and the dimension of the vector is the same as the number of image feature factors.

[0048] The initial feature vectors corresponding to each patch in the image are stored sequentially to construct the initial feature set of the image, denoted as .

[0049] H (0) = [f1, f2, ..., fn ] T ;

[0050] Among them, H (0) This serves as the initial feature set for a single image, used as the input layer for subsequent functional area recognition models. d is the image feature vector of the i-th plot; input The feature dimension of the initial feature vector is the same as the number of factors of the image feature factors.

[0051] It should be noted that, in order to eliminate the influence caused by the different dimensions and numerical ranges of the feature factors, in some embodiments, it is necessary to standardize the feature values ​​of different image feature factors in the image; for example, Z-score standardization is used to make the feature values ​​of different image feature factors have a mean of 0 and a standard deviation of 1.

[0052] S300: For each patch in the image, perform spatial proximity analysis and image feature similarity analysis to construct the spatial proximity map and semantic similarity map corresponding to the image.

[0053] Specifically, the method for constructing the spatial proximity graph includes:

[0054] S301, extract the center point of each patch and set it as a node in the spatial proximity map;

[0055] The center point is a point used to characterize the spatial location of a land parcel; for example, the center point can be a center point, the center point of a circumscribed rectangle, the center point of an inscribed circle, etc.

[0056] Nodes are objects used to represent land parcels in the spatial proximity map, and there is a one-to-one correspondence between nodes and land parcels, that is, a single land parcel corresponds to a single node.

[0057] Specifically, the geometric center point of each plot of land is extracted from the image and used as a node in the spatial proximity map, as follows:

[0058] V = {v1, v2, ..., v} n}

[0059] Where V is the set of nodes, v i Represents the plot unit s i The nodes corresponding to the center point, and each center point v i It has corresponding spatial coordinates (x, y); n is the total number of nodes in the spatial proximity map, which is the same as the total number of land parcels in the image.

[0060] S302, Based on the spatial location of each node in the image, use geometric graphics methods to calculate the degree of adjacency between the spatial locations of each node, and obtain the spatial neighbor nodes corresponding to the node;

[0061] Specifically, each node in the image is constructed into a node set V of the image;

[0062] Based on the spatial distribution of each node, standard Delaunay triangulation is performed on each node to form a continuous Delaunay triangulation network. For any triangle in this network, its three corner points (nodes) are spatially adjacent to each other, that is, the land parcels corresponding to each corner point (node) in a single triangle are adjacent to each other. Furthermore, the circumcircle of any triangle does not contain any other nodes in the Delaunay triangulation network, thus enabling the automatic and accurate extraction of adjacency relationships between nodes.

[0063] The triangles in the Delaunay triangulation are aggregated into a single set, which is then output as:

[0064] T = {t1, t2, ..., t} n};

[0065] Where T is the total set of triangles, t i For any triangle in the Delaunay triangulation, it is defined by the index (m,n,k) of its corresponding center point; that is, t i Center point v m ,v n ,v k The triangle formed.

[0066] For a single node, extract other nodes that are spatially adjacent to the current node from the triangle set and use them as the spatial neighbors of the current node; add all the spatial neighbors of the current node to a set to construct the spatial neighbor set of the current node; traverse each node to obtain the spatial neighbor set of each node.

[0067] S303, construct the edges between a node and its corresponding spatial neighbor nodes, and set them as edges in the spatial proximity graph;

[0068] For each other node in the spatial neighbor set, create an undirected edge between the current node and other nodes, and add the edge to the preset edge set;

[0069] Each extracted edge is deduplicated to construct an edge set of the image and ensure that each edge is unique within that edge set.

[0070] S304, based on the set of nodes and the set of edges, constructs the spatial proximity graph corresponding to the image, as follows:

[0071] G spatial =(V,E) spatial );

[0072] In the formula, G spatial Let V be a spatial proximity graph, which is an undirected graph; V is the set of nodes; E is a spatial proximity graph. spatial Let it be the set of edges.

[0073] After obtaining the spatial proximity map corresponding to the image, the spatial adjacency matrix corresponding to the spatial proximity map is extracted as follows:

[0074]

[0075] In the formula, A spatial Let A be the spatial adjacency matrix, used to represent the spatial adjacency information between nodes in the spatial proximity graph, i.e., whether they are spatially adjacent or not; i and j are two nodes in the spatial proximity graph; A spatial [i,j]=1 indicates that the land parcel s corresponding to node i is... i The land parcel s corresponding to node j j They are directly adjacent in space.

[0076] This application extracts the spatial adjacency relationships between nodes through Delaunay triangulation, thereby directly converting the spatial proximity relationships between various patch features in the image into processable matrix information.

[0077] In one optional embodiment, the semantic similarity graph is constructed in the following ways:

[0078] S310, set the center point of each patch as a node in the semantic similarity graph;

[0079] In this context, nodes are objects used in the semantic similarity graph to represent the corresponding land parcels, and there is a one-to-one correspondence between nodes and land parcels, that is, a single land parcel corresponds to a single node.

[0080] Specifically, the geometric center point of each land parcel in the image is extracted and used as a node in the semantic similarity graph, as follows:

[0081] V = {v1, v2, ..., v} n}

[0082] Where V is the set of nodes; n is the total number of nodes in the semantic similarity graph, which is the same as the total number of land parcels in the image; v i Represents the plot unit s i The nodes corresponding to the center point, and each center point v i It has corresponding spatial coordinates (x, y).

[0083] S320: Based on the image feature vectors of each node in the image, feature similarity calculation is used to extract the similarity between the image feature vectors corresponding to each node and obtain the semantic neighbor nodes corresponding to the nodes.

[0084] Specifically, the image feature vector of a land parcel is set as the image feature vector of the node corresponding to that land parcel, that is, each node is a d-dimensional real vector; where d is the dimension of the image feature and is the same as the number of image feature factors.

[0085] For a single node, iterate through all other nodes and calculate the similarity between the current node and other nodes on the image feature vector, which is used as the semantic similarity between the two nodes. For example, the cosine similarity formula is used to calculate the similarity between two nodes on the image feature vector.

[0086] The semantic similarity scores between the current node and all other nodes are sorted according to their scores, resulting in a sorted list for the current node. Using k-nearest neighbor selection, the k nodes with the highest scores are chosen from this sorted list as the nodes most similar to the current node in terms of image features, i.e., semantic neighbors. All semantic neighbors corresponding to the current node are added to a set to construct the semantic neighbor set for the current node. Based on this step, each node is traversed to obtain the semantic neighbor set for each node. The value of k is set empirically to balance the connectivity of the structure graph and the computational complexity.

[0087] S330, construct the edges between the node and its corresponding semantic neighbor nodes, and set them as edges in the semantic similarity graph;

[0088] For each other node in the semantic neighbor set, establish an undirected edge (v) between the current node and other nodes. i ,v j And add the edge to the edge set E. semantic .

[0089] It should be noted that, in other implementations, the similarity between nodes can also be obtained by calculating the reciprocal of the Euclidean distance.

[0090] In an optional embodiment, since the k-nearest neighbor relationship is asymmetric (i.e., v j It is v i k nearest neighbors, but v i It's not necessarily v j The k-nearest neighbors (E) are needed, therefore, after traversing all nodes, the edge set E needs to be checked. semantic Each edge in the graph is symmetrically processed to ensure the connectivity and consistency of relationships in the graph.

[0091] In a more specific embodiment, the symmetry processing includes:

[0092] Perform union symmetry on two overlapping edges, i.e., for the first edge (v... i ,v j) and the second side (v) that coincides with the first side. j ,v i If the first side (v) i ,v j )∈E edge set E semantic Or the second side (v) j ,v i )∈E edge set E semantic Then in the final edge set E semantic The first side (v) is retained in the middle. i ,v j ).

[0093] In another, more specific embodiment, the symmetry processing includes:

[0094] Perform intersection symmetry on two overlapping edges, i.e., for the first edge (v... i ,v j ) and the second side (v) that coincides with the first side. j ,v i ), only if the first side (v i ,v j )∈E edge set E semantic And the second side (v j ,v i )∈E edge set E semantic Then in the final edge set E semantic The first side (v) is retained in the middle. i ,v j ).

[0095] S340, based on the set of nodes and the set of edges, constructs a semantic similarity graph corresponding to the image, as follows:

[0096] G semantic =(V,E) semantic );

[0097] Among them, G semantic Let V be a semantic similarity graph, which is an undirected graph; V is the set of nodes; E is the set of nodes. semantic Let it be the set of edges.

[0098] After obtaining the semantic similarity map corresponding to the image, the semantic adjacency matrix of the semantic similarity map is obtained as follows:

[0099]

[0100] In the formula, A semantic Let A be the semantic adjacency matrix, used to represent the semantic similarity information between nodes in the semantic similarity graph, i.e., whether they are semantically similar or not; i and j are two nodes in the semantic similarity graph; A semantic [i, j] = 1 indicates that the land parcel s corresponding to node i is... iThe land parcel s corresponding to node j j Semantic similarity.

[0101] S400 constructs sample data corresponding to the image based on the initial feature set, spatial proximity map, semantic similarity map and functional area information of the image;

[0102] Specifically, for a single scene image, functional area patches corresponding to the image are obtained; the functional area patches include the spatial distribution and functional area type of the patch; based on the spatial location of the plot patch and the spatial distribution of the functional area patch, the two are spatially overlaid to obtain the functional area type corresponding to each plot patch, so as to obtain the functional area distribution corresponding to the image.

[0103] For each image, the initial feature set, spatial proximity map, and semantic similarity map corresponding to the image are integrated, and the integrated data is used as the feature input of the sample data; the functional area distribution corresponding to the image is used as the label of the sample data to construct the sample data corresponding to the image.

[0104] S500, based on sample data, trains a preset functional area recognition model to obtain a trained functional area recognition model;

[0105] The functional area identification model is a hierarchical graph neural network model constructed based on spatial proximity graphs and semantic similarity graphs, and includes at least a first graph neural network layer for performing spatial proximity enhancement and a second graph neural network layer for performing semantic similarity enhancement.

[0106] Specifically, the functional area recognition model includes: an input layer, a first graph neural network layer, a second graph neural network layer, a feature fusion layer, and an output layer;

[0107] The input layer is used for the initial feature set, the spatial proximity map, the semantic similarity map, and the barrier map structure of the input image;

[0108] The first neural network layer connects to the input layer and is used to obtain the spatial proximity map and initial feature set input through the input layer. It then performs enhancement processing on the initial feature set based on the spatial proximity map to obtain a first intermediate feature distribution that incorporates spatial context information.

[0109] The second neural network layer connects the input layer and the first neural network layer respectively. It is used to obtain the semantic similarity map input by the input layer and the first intermediate feature distribution output by the first neural network layer. Based on the semantic similarity map, feature enhancement processing is performed on the first intermediate feature distribution to obtain the second intermediate feature distribution that incorporates the global semantic context.

[0110] The feature fusion layer connects the first graph neural network layer and the second graph neural network layer, and is used to fuse the features in the first intermediate feature distribution and the second intermediate feature distribution to obtain the fused feature distribution.

[0111] The output layer is connected to the feature fusion layer, which is used to perform classification prediction based on the fused feature distribution and the classifier to obtain the recognition prediction results of the functional area.

[0112] It should be noted that both the first graph neural network layer and the second graph neural network layer are any type of network layer capable of enabling information transfer between nodes, including but not limited to graph convolutional network layers and graph attention network layers.

[0113] In this application, each sample data is input into a preset model to obtain the functional area prediction results corresponding to each sample data; based on the prediction results and sample labels (actual distribution of functional areas), the trainable parameters in the model are optimized using a loss function to realize the training process of the model.

[0114] The loss function is the cross-entropy loss function; the trainable parameters are obtained by gradient descent algorithm (such as Adam optimizer), and the gradient is calculated through backpropagation to update the parameters until the model training reaches convergence or meets the exit condition.

[0115] S600 uses the trained functional area recognition model to perform the functional area recognition process of the image to be recognized and obtain the corresponding functional area recognition results.

[0116] Specifically, the process involves acquiring images of the functional regions to be identified; and then using the trained functional region recognition model to perform the following steps on the images:

[0117] Based on preset image feature factors, image feature vectors of each patch in the image to be identified are extracted to obtain the initial feature set corresponding to the image to be identified; and, by performing spatial proximity analysis and image feature similarity analysis on each patch, spatial proximity map and semantic similarity map corresponding to the image to be identified are constructed respectively.

[0118] The initial feature set, spatial proximity map, and semantic similarity map corresponding to the image to be identified are integrated, and the integrated data is input into the trained functional area recognition model to obtain the functional area recognition result corresponding to the image; wherein, the functional area recognition result includes the type and distribution range of the functional area.

[0119] In one optional embodiment, the graph neural network layer is a graph convolutional layer; that is, the first graph neural network layer is a spatial graph convolutional layer, and the second graph neural network layer is a semantic graph convolutional layer.

[0120] During model training, the spatial graph-based convolutional layers perform the following:

[0121] 1) Add a self-join to the spatial adjacency matrix, as follows:

[0122]

[0123] In the formula, For the spatial adjacency matrix after adding self-joins; A spatial Let I be the spatial adjacency matrix; let I be the identity matrix, satisfying Where n is the number of nodes in the spatial adjacency graph, which is the same as the number of land parcels in the image.

[0124] 2) Calculate the degree matrix of the spatial adjacency matrix; the degree matrix of the spatial adjacency matrix is ​​calculated using existing degree matrix calculation methods, which will not be elaborated here.

[0125] 3) Perform symmetric normalized graph convolution on the initial feature set based on the degree matrix to obtain the first intermediate feature distribution, which is:

[0126]

[0127] In the formula, The first intermediate feature distribution satisfies n is the total number of nodes in the current spatial adjacency graph; H (0) This is the initial feature set; for The degree matrix; Let be a trainable weight matrix in a spatial graph convolutional layer, which satisfies ... d input Let d be the initial feature dimension of each node. hidden1 σ represents the feature dimension output by each node in the spatial graph convolutional layer; σ is the activation function, preferably the ReLU activation function.

[0128] Similarly, during model training, the following steps are performed based on the semantic graph convolutional layer:

[0129] 1) Add self-connects to the semantic adjacency matrix, as follows:

[0130]

[0131] In the formula, The semantic adjacency matrix after adding self-connections; A semantic Let I be the semantic adjacency matrix; let I be the identity matrix, satisfying Where n is the number of nodes in the semantic similarity graph, which is the same as the number of land parcels in the image.

[0132] 2) Calculate the degree matrix of the semantic adjacency matrix. The degree matrix of the semantic adjacency matrix is ​​calculated using the existing degree matrix calculation method, which will not be elaborated here.

[0133] 3) Perform symmetric normalized graph convolution on the first intermediate feature distribution based on the degree matrix to obtain the second intermediate feature distribution, which is:

[0134]

[0135] In the formula, The second intermediate feature distribution satisfies for The degree matrix; Let be a trainable weight matrix for the semantic graph convolutional layer, and satisfy . d hidden1 d represents the feature dimension output by each node in the spatial graph convolutional layer; hidden2 σ represents the feature dimension output by each node in the semantic graph convolutional layer; σ is the activation function, preferably the ReLU activation function.

[0136] In another optional embodiment, to improve the model's ability to distinguish important neighbor nodes, the graph neural network layer adopts a graph attention network layer; that is, the first graph neural network layer adopts a spatial graph attention layer, and the second graph neural network layer adopts a semantic graph attention layer, so that each node can adaptively and discriminatively obtain local spatial context information and semantic context information related to the functional area from its corresponding neighbor nodes.

[0137] Specifically, during model training, the spatial graph attention layer performs the following:

[0138] 1) Perform a linear transformation on the initial feature set under different attention mechanisms to obtain the initial feature distribution after the linear transformation of each attention mechanism;

[0139] The initial feature distribution after linear transformation corresponds to the attention mechanism; that is, under different attention mechanisms, the initial feature distribution after linear transformation corresponding to the attention mechanism is obtained respectively.

[0140] Specifically, under the k-th attention mechanism, a linear transformation is performed on the initial feature set of the input, as follows:

[0141]

[0142] In the formula, Z (0),k Let the initial feature distribution after linear transformation under the k-th attention mechanism satisfy the following conditions: Where n is the number of nodes in the spatial adjacency graph, which is the same as the total number of land parcels in the image; d head1The feature dimension of the node in the spatial attention layer; Let be the trainable weight matrix of the spatial graph attention layer, which satisfies: d input The feature dimension of the node in the initial feature set.

[0143] 2) For each node in the spatial adjacency graph, construct the first intermediate feature vector for each node based on the feature vectors of the node and its corresponding spatial neighbor nodes in the initial feature distribution after linear transformation.

[0144] Specifically, for a single node in the spatial adjacency graph, this step, when executed, includes:

[0145] A. Obtain all spatial neighbor nodes of the current node in the spatial proximity graph;

[0146] B. For each spatial neighbor node, based on the feature vectors of the current node and spatial neighbor nodes in the initial feature distribution after linear transformation (under the k-th attention mechanism), construct the spatial attention coefficient between them, which serves as the spatial attention coefficient between the current node and spatial neighbor nodes (under the k-th attention mechanism), as follows:

[0147]

[0148] In the formula, Let be the spatial attention coefficient between the current node i and its spatial neighbor node j under the k-th attention mechanism; Let be the attention vector under the k-th attention mechanism, and be the learnable attention vector; || denotes the vector concatenation operation; σ is the activation function, preferably the LeakyReLU activation function.

[0149] C. Normalize the spatial attention coefficients (under the k-th attention mechanism) corresponding to each spatial neighbor node as follows:

[0150]

[0151] In the formula, Let A be the normalized spatial attention coefficient between the current node i and its spatial neighbor node j under the k-th attention mechanism; l∈{l|A spatial [i, l] = 1} represents each spatial neighbor of node i.

[0152] D. Based on the normalized spatial attention coefficients between the current node and its neighboring nodes (under the k-th attention mechanism), the initial feature vectors of each neighboring node after linear transformation (under the k-th attention mechanism) are weighted, summed, and activated to obtain the first intermediate feature vector corresponding to the current node (under the k-th attention mechanism); as follows:

[0153]

[0154] In the formula, σ is the activation function, preferably the ELU activation function.

[0155] E. Concatenate the first intermediate feature vectors of the current node under each attention mechanism to obtain the first intermediate feature vector corresponding to the current node, as follows:

[0156]

[0157] In the formula, Wherein, K1 represents the total number of attention mechanisms in the spatial graph attention layer.

[0158] Perform steps A through E above on each node in the spatial adjacency graph to obtain the first intermediate feature vector corresponding to each node.

[0159] 3) Merge the first intermediate feature vectors corresponding to each node to obtain the first intermediate feature distribution corresponding to the sample data.

[0160] Wherein, the first intermediate feature distribution satisfies

[0161] Similarly, during model training, the following steps are performed based on the semantic graph attention layer:

[0162] 1) Under different attention mechanisms, perform a linear transformation on the first intermediate feature distribution to obtain the first intermediate feature distribution after linear transformation with each attention mechanism;

[0163] The first intermediate feature distribution after linear transformation corresponds to the attention mechanism; that is, under different attention mechanisms, the first intermediate feature distribution after linear transformation corresponding to the attention mechanism is obtained respectively.

[0164] Specifically, under the k-th attention mechanism, a linear transformation is performed on the first intermediate feature distribution of the input, as follows:

[0165]

[0166] In the formula, Z (1)，k Let be the first intermediate feature distribution after linear transformation under the k-th attention mechanism, satisfying Where n is the total number of nodes in the current semantic similarity graph, which is the same as the total number of land parcels in the image; d head2 The feature dimension of the node in the semantic graph attention layer; Let be the weight matrix of the semantic graph attention layer under the k-th attention mechanism, and let be a trainable weight matrix that satisfies: Wherein K1×d head1The first intermediate feature distribution The feature dimensions of each node.

[0167] 2) For each node in the semantic similarity graph, construct the second intermediate feature vector corresponding to each node based on the feature vectors of the node and its corresponding semantic neighbor nodes in the first intermediate feature distribution after linear transformation.

[0168] Specifically, for a single node in the semantic similarity graph, this step, when executed, includes:

[0169] A. Retrieve all semantic neighbor nodes of the current node;

[0170] B. For each semantic neighbor node, based on the feature vectors of the current node and the semantic neighbor nodes in the first intermediate feature distribution after linear transformation (under the k-th attention mechanism), construct the semantic attention coefficient between them, which serves as the semantic attention coefficient between the current node and the neighbor node (under the k-th attention mechanism), as follows:

[0171]

[0172] In the formula, Let be the semantic attention coefficient between the current node i and its semantic neighbor node j under the k-th attention mechanism; and These are the feature vectors of the current node i and its semantic neighbor node j in the first intermediate feature distribution after linear transformation (under the k-th head attention mechanism); Let be the attention vector under the k-th attention mechanism, and be the learnable attention vector; || denotes the vector concatenation operation; σ is the activation function, preferably the LeakyReLU activation function.

[0173] C. Normalize the semantic attention coefficients (under the k-th attention mechanism) corresponding to each neighbor node, as follows:

[0174]

[0175] In the formula, Let I represent the normalized semantic attention coefficients between the current node i and its semantic neighbor node j under the k-th attention mechanism; I∈{l|A semantic [i, l] = 1} represents each semantic neighbor of node i.

[0176] D. Based on the normalized semantic attention coefficients between the current node i and its semantic neighbor node j (under the k-th attention mechanism), the first intermediate feature vectors (under the k-th attention mechanism) after linear transformation of each neighbor node are weighted, summed, and activated to obtain the second intermediate feature vector (under the k-th attention mechanism) corresponding to the current node, which is:

[0177]

[0178] Where σ is the activation function, preferably the ELU activation function.

[0179] E. Concatenate the second intermediate feature vectors of the current node under each attention mechanism to obtain the second intermediate feature vector corresponding to the current node, as follows:

[0180]

[0181] in, Wherein, K2 represents the total number of attention mechanisms in the semantic graph attention layer.

[0182] Perform steps A through E above on each node in the semantic similarity graph to obtain the second intermediate feature vector corresponding to each node.

[0183] 3) Merge the second intermediate feature vectors corresponding to each node to obtain the second intermediate feature distribution corresponding to the sample data.

[0184] in, satisfy:

[0185] To comprehensively utilize the contextual information in the spatial proximity graph and the semantic similarity graph, a feature fusion layer is used to fuse the feature distributions corresponding to the same node in the outputs of the two graph neural network layers mentioned above, obtaining the fused feature distribution of the sample data, as follows:

[0186]

[0187] in, d fused This represents the total feature dimension after concatenation.

[0188] Based on the fusion feature distribution of nodes, combined with the functional area type information corresponding to the nodes, the output layer is used to perform functional area classification prediction on each node to obtain the functional area prediction results corresponding to each node.

[0189] Specifically, the steps include:

[0190] a) Linear transformation: Through a fully connected layer, the high-dimensional features in the fused feature distribution are mapped to a dimension equal to the number of types C of the functional area.

[0191]

[0192] in, It is a trainable weight matrix. is the original classification score matrix of all nodes; C is the number of functional area types.

[0193] b) Probability Normalization: Z is normalized using the Softmax function to obtain the class probability vector of each node.

[0194] c) Category prediction: Take the functional area type corresponding to the maximum value in the category probability vector of each node as its final predicted label.

[0195]

[0196] The method provided in this application adopts a technical logic chain of "superpixel segmentation -> multi-dimensional feature extraction -> multi-level graph construction -> hierarchical graph neural network training and inference". It realizes the deep mining of the correlation features between various patches in the image within the functional area, and uses graph neural networks to perform deep learning on the graph structure topological relationship based on spatial proximity and semantic similarity. This not only reduces the dependence on multi-source auxiliary data such as POI and street view, but also greatly reduces the amount of sample data used for model training. That is, it can achieve end-to-end deep learning model training based only on remote sensing image and a small amount of sample data, and obtain a functional area recognition model with good recognition effect. In addition, the method of this application greatly reduces the tedious manual parameter tuning in traditional model training methods, and effectively promotes the engineering application of the technology.

[0197] Because the same type of geographical feature within an urban area may be blocked by geographical barriers such as rivers and roads; for example, such as Figure 3 As shown, when a road (the yellow straight line in the figure) passes through a residential area, the area is divided into a first building sub-area (the area marked by the red dashed line in the figure) and a second building sub-area (the area marked by the orange dashed line in the figure). When the image features of the houses in the two sub-areas are similar (the image features of the roofs of the houses within the red solid line border and the roofs within the orange solid line border are similar), based on existing identification methods, the first building sub-area located on the left side of the road and the second building sub-area located on the right side are often identified as the same functional area. However, in the actual distribution of functional areas, the first building sub-area belongs to the educational area, while the second building sub-area belongs to the residential area.

[0198] Based on this, it can be seen that in existing technologies, the impact of geographical barriers on the division of functional zones is often overlooked, which leads to geographical elements located on different sides of geographical barriers being misidentified as the same functional zone (while in actual planning work, geographical elements located on different sides of geographical barriers are usually divided into different functional zones). For example, the first building sub-zone located on one side of the road and the second building sub-zone located on the other side are identified as the same functional zone, which leads to a decrease in the accuracy of functional zone identification.

[0199] To enhance the constraints of geographical barriers on the functional zone identification process and further improve the accuracy of the functional zone identification model, this application provides another functional zone identification method for identifying functional zone information within an urban area (where geographical barriers exist).

[0200] Please see Figure 4 The diagram illustrates a flowchart of a functional area identification method in an embodiment, which includes the following steps:

[0201] S10, acquire remote sensing images covering the functional area, wherein the remote sensing images contain geographical barrier elements;

[0202] In this embodiment, the remote sensing image contains geographic barrier features, so each segmented patch also contains patches corresponding to the geographic barrier features; for example, patches corresponding to main roads or rivers. Based on this, to avoid the influence of patches corresponding to geographic barrier features on subsequent steps, after segmenting the remote sensing image using an object-oriented approach to obtain each segmented patch, patches overlapping with geographic barrier features are removed from each segmented patch, and the remaining patches are treated as land parcels for processing in subsequent steps.

[0203] The land parcels and geographic elements are in a one-to-one correspondence, meaning that each land parcel corresponds to one geographic element.

[0204] Specifically, the acquisition method for each patch is the same as that in the above embodiments, and will not be repeated here.

[0205] It should be noted that in some other embodiments, the geographic barrier element is a line element. In this case, after obtaining each segmented patch, it is not necessary to remove the patches that overlap with the geographic barrier element to conveniently perform subsequent steps.

[0206] S20, based on preset image feature factors, extract image feature vectors corresponding to each patch in the image to construct an initial feature set corresponding to the image;

[0207] Among them, image feature factors are used to characterize the feature information of land parcels on images, including one or more of spectral feature factors, texture feature factors, and morphological feature factors.

[0208] In this embodiment, the initial feature set is constructed in the same way as in the above embodiments, and will not be described again here.

[0209] S30, For each patch in the image, perform spatial proximity analysis, image feature similarity analysis and geographic barrier analysis respectively, so as to construct the spatial proximity map, semantic similarity map and barrier map structure corresponding to the image.

[0210] The spatial proximity graph and the semantic similarity graph are constructed in the same way as in the above embodiments, and will not be described again here.

[0211] In this application, the barrier map structure is constructed in the following ways:

[0212] S31, extract the center point of each patch and set it as a node in the barrier map structure;

[0213] The center point is a point used to characterize the spatial location of a land parcel; for example, the center point can be a center point, the center point of a circumscribed rectangle, the center point of an inscribed circle, etc.

[0214] Nodes are objects in the barrier graph structure used to represent the corresponding land parcels. There is a one-to-one correspondence between nodes and land parcels, that is, a single land parcel corresponds to a single node.

[0215] Specifically, the geometric center point of each plot of land is extracted from the image and used as a node in the barrier map structure, as follows:

[0216] V = {v1, v2, ..., v} n}

[0217] Where V is the set of nodes, v i Represents the plot unit s i The nodes corresponding to the center point, and each center point v i It has corresponding spatial coordinates (x, y); n is the total number of nodes in the barrier map structure, which is the same as the total number of land parcels in the image.

[0218] S32, Based on the spatial location of each node in the image and the spatial location of the geographic barrier element, obtain the barrier neighbor node corresponding to the node by determining the position between any two nodes and the geographic barrier element.

[0219] Specifically, initialize an empty set of edges, denoted as E. barrier ={};

[0220] For a single node v i According to the current node v i Other nodes v j The spatial location of (i≠j) is determined by combining the spatial location of geographical barriers; the current node v is then determined. i Other nodes v j Whether they are located on different sides of the same geographical barrier element, if so, then the land parcels corresponding to the two nodes are considered to be blocked by that geographical barrier element, i.e., other nodes v j As the current node v iThe blocking neighbor nodes; add all the blocking neighbor nodes corresponding to the current node to a set to construct the blocking neighbor set of the current node;

[0221] Based on this step, each node is traversed to obtain the set of blocking neighbors corresponding to each node.

[0222] S33, construct the edge between the node and its corresponding blocking neighbor node, and set it as the edge in the blocking graph structure;

[0223] For each blocking neighbor node in the blocking neighbor set, establish an undirected edge (v_blocking_neighbors) between the current node and the blocking neighbor node. i ,v j And add the edge to the edge set E. barrier .

[0224] S34, based on the node set and edge set, constructs the barrier graph structure corresponding to the image, as follows:

[0225] G barrier =(V,E) barrier );

[0226] Among them, G barrier It is a barrier graph structure, which is an undirected graph; V is the set of nodes; E barrier Let it be the set of edges.

[0227] After obtaining the barrier map structure corresponding to the image, the barrier adjacency matrix of this barrier map structure is obtained as follows:

[0228]

[0229] In the formula, A barrier Let A be the barrier adjacency matrix, used to characterize the barrier information between nodes in the barrier graph structure, i.e., whether they are blocked or not. Let i and j be two nodes in the barrier graph structure; A barrier [i,j]=1 indicates that the land parcel s corresponding to node i is... i The land parcel s corresponding to node j j There is a barrier between them.

[0230] S40, based on the initial feature set, spatial proximity map, semantic similarity map, barrier structure map and functional area information of each image, construct sample data;

[0231] Specifically, for a single scene image, functional area patches corresponding to the image are obtained; the functional area patches include the spatial distribution and functional area type of the patch; based on the spatial location of the plot patch and the spatial distribution of the functional area patch, the two are spatially overlaid to obtain the functional area type corresponding to each plot patch, so as to obtain the functional area distribution corresponding to the image.

[0232] For each image, the initial feature set, spatial proximity map, semantic similarity map, and barrier structure map corresponding to the image are integrated, and the integrated data is used as the feature input of the sample data; the functional area distribution corresponding to the image is used as the label of the sample data to construct the sample data corresponding to the image.

[0233] S50, based on sample data, train the preset functional area recognition model to obtain the trained functional area recognition model;

[0234] The functional area identification model is a hierarchical graph neural network model constructed based on the structure of spatial proximity graph, semantic similarity graph and barrier graph, and includes at least: a first graph neural network layer for performing spatial proximity and geographical barrier enhancement, and a second graph neural network layer for performing semantic similarity and geographical barrier enhancement.

[0235] Specifically, the functional area recognition model includes: an input layer, a first graph neural network layer, a second graph neural network layer, a feature fusion layer, and an output layer;

[0236] The input layer is used for the initial feature set, the spatial proximity map, the semantic similarity map, and the barrier map structure of the input image;

[0237] The first neural network layer connects to the input layer and is used to obtain the initial feature set, spatial proximity map, and barrier map structure input through the input layer. It then performs enhancement processing on the initial feature set based on the spatial proximity map and barrier map structure to obtain a first intermediate feature distribution that integrates spatial context information and geographical barrier information.

[0238] The second neural network layer connects the input layer and the first neural network layer respectively, and is used to obtain the semantic similarity map and barrier map structure input by the input layer, as well as the first intermediate feature distribution output by the first neural network layer; based on the semantic similarity map and barrier map structure, feature enhancement processing is performed on the first intermediate feature distribution to obtain a second intermediate feature distribution that incorporates global semantic context information and geographical barrier information;

[0239] The feature fusion layer connects the first graph neural network layer and the second graph neural network layer, and is used to fuse the features in the first intermediate feature distribution and the second intermediate feature distribution to obtain the fused feature distribution.

[0240] The output layer is connected to the feature fusion layer, which is used to perform classification prediction based on the fused feature distribution and the classifier to obtain the recognition prediction results of the functional area.

[0241] It should be noted that both the first graph neural network layer and the second graph neural network layer are any type of network layer capable of enabling information transfer between nodes, including but not limited to graph convolutional network layers and graph attention network layers.

[0242] In this application, sample data is input into a preset model to obtain the functional area prediction results corresponding to each sample data. Based on the prediction results and sample labels (actual information of the functional areas), the trainable parameters in the model are optimized using a loss function to achieve the model training process. The loss function used is the cross-entropy loss function; the trainable parameters are updated using a gradient descent algorithm (such as the Adam optimizer) through backpropagation to calculate gradients until the model training reaches convergence or meets the exit conditions.

[0243] S60: Acquire the image to be identified and extract the geographical barrier features in the image; use the trained functional area identification model to perform the functional area identification process of the image to be identified and obtain the corresponding functional area identification results.

[0244] Specifically, the process involves acquiring an image of the functional area to be identified and extracting geographic barrier features from that image; and performing the following steps on the image to be identified:

[0245] Based on the preset image feature factors, the image feature vectors of each patch in the image to be identified are extracted to obtain the initial feature set corresponding to the image to be identified; and, by performing spatial proximity analysis, image feature similarity analysis and geographical barrier analysis on each patch, the spatial proximity map, semantic similarity map and barrier map structure corresponding to the image to be identified are constructed respectively.

[0246] The initial feature set, spatial proximity map, semantic similarity map, and barrier map structure corresponding to the image to be identified are integrated, and the integrated data is input into the trained functional area recognition model to obtain the functional area recognition result corresponding to the image; wherein, the functional area recognition result includes at least the type and distribution range of the functional area.

[0247] In one optional embodiment, the graph neural network layer is a graph convolutional layer; that is, the first graph neural network layer is a spatial graph convolutional layer, and the second graph neural network layer is a semantic graph convolutional layer.

[0248] During model training, the spatial graph-based convolutional layers perform the following:

[0249] 1) Perform deblocking processing on the spatial neighbor matrix based on the blocking adjacency matrix to obtain a new spatial neighbor matrix;

[0250] Specifically, based on the geographical barriers between nodes in the adjacency matrix, edges with geographical barriers are removed from the spatial proximity matrix, and the resulting matrix is ​​used as the new spatial proximity matrix:

[0251] A′ spatial =A spatial ☉(1-A barrier)

[0252] In the formula, A′ spatial For the new spatial proximity matrix; A spatial A is the spatial nearest neighbor matrix before deblocking; barrier For blocking adjacency matrices; ⊙ Element-by-element multiplication;

[0253] 2) Add a self-connect matrix to the new spatial adjacency matrix, as follows:

[0254]

[0255] Where I is the identity matrix, satisfying Where n is the number of nodes in the spatial adjacency graph, which is the same as the number of land parcels in the image.

[0256] 3) Calculate the degree matrix of the new spatial adjacency matrix:

[0257] The degree matrix of the new spatial adjacency matrix is ​​calculated using the existing degree matrix calculation method, which will not be elaborated here.

[0258] 4) Based on the degree matrix, perform symmetric normalized graph convolution on the initial feature set to obtain the first intermediate feature distribution, which is:

[0259]

[0260] In the formula, The first intermediate feature distribution satisfies n is the total number of nodes in the current spatial adjacency graph; for The degree matrix; Let be a trainable weight matrix in a spatial graph convolutional layer, which satisfies ... d input Let d be the initial feature dimension of each node. hidden1 σ represents the feature dimension output by each node in the spatial graph convolutional layer; σ is the activation function, preferably the ReLU activation function.

[0261] Similarly, the second graph neural network layer employs a semantic graph convolutional layer; during model training, the following operations are performed based on the semantic graph convolutional layer:

[0262] 1) Based on the blocking adjacency matrix, perform deblocking processing on the semantic adjacency matrix to obtain a new semantic adjacency matrix;

[0263] A′ semantic =A semantic ☉(1-A barrier )

[0264] In the formula, A′ semantic For the new semantic adjacency matrix; Asemantic A is the semantic adjacency matrix before deblocking; barrier To isolate the adjacent matrix; ⊙ Element-by-element multiplication.

[0265] 2) Add a self-connect matrix to the new semantic adjacency matrix, for

[0266]

[0267] Where I is the identity matrix, satisfying Where n is the number of nodes in the semantic similarity graph, which is the same as the number of land parcels in the image.

[0268] 3) Calculate the degree matrix of the new semantic adjacency matrix;

[0269] The degree matrix of the new semantic adjacency matrix is ​​calculated using the existing degree matrix calculation method, which will not be elaborated here.

[0270] 4) Perform symmetric normalized graph convolution on the first intermediate feature distribution based on the degree matrix to obtain the second intermediate feature distribution, which is:

[0271]

[0272] in, The second intermediate feature distribution satisfies for The degree matrix; Let be a trainable weight matrix for the semantic graph convolutional layer, and satisfy . d hidden1 d represents the feature dimension output by each node in the spatial graph convolutional layer; hidden2 σ represents the feature dimension output by each node in the semantic graph convolutional layer; σ is the activation function, preferably the ReLU activation function.

[0273] In another optional embodiment, to improve the model's ability to distinguish important neighbor nodes, the graph neural network layer adopts a graph attention network layer; that is, the first graph neural network layer is a spatial graph attention layer with geographical barrier awareness, which enables each node to adaptively and discriminatively obtain local spatial context information related to the functional area and not blocked from its corresponding neighbor nodes based on geographical constraints; and the second graph neural network layer is a semantic graph attention layer with geographical barrier awareness, which enables each node to adaptively and discriminatively obtain semantic context information related to the functional area and not blocked from its corresponding neighbor nodes based on geographical constraints.

[0274] Specifically, during model training, the spatial graph attention layer performs the following:

[0275] 1) Obtain the initial feature distribution of the initial feature set after linear transformation under different attention mechanisms;

[0276] Under different attention mechanisms, a linear transformation is performed on the initial feature set to obtain the initial feature distribution after linear transformation corresponding to each attention mechanism; wherein, the initial feature distribution after linear transformation corresponds to the attention mechanism; that is, under different attention mechanisms, the initial feature distribution after linear transformation corresponding to the attention mechanism is obtained respectively.

[0277] Specifically, under the k-th attention mechanism, a linear transformation is performed on the initial feature set of the input, as follows:

[0278]

[0279] In the formula, Z (0),k Let the initial feature distribution after linear transformation under the k-th attention mechanism satisfy the following conditions: Where n is the number of nodes in the spatial adjacency graph, which is the same as the total number of land parcels in the image; d head1 The feature dimension of the node in the spatial attention layer; Let be the trainable weight matrix in the spatial graph attention layer, which satisfies: d input The feature dimension of the node in the initial feature set.

[0280] 2) Based on the feature vectors of the node and its corresponding spatial neighbor nodes in the initial feature distribution after linear transformation, and combined with the barrier information corresponding to the node, construct the spatial attention coefficient between the two, so as to enhance the attention of the node's initial feature vector after linear transformation based on the spatial attention coefficient, and obtain the first intermediate feature distribution under the comprehensive geographical barrier information.

[0281] Specifically, for each node in the spatial proximity graph, a first intermediate feature vector of the node is constructed based on the comprehensive barrier information, including:

[0282] A. Obtain all spatial neighbor nodes of the current node in the spatial proximity graph;

[0283] B. For each spatial neighbor node, based on the feature vectors of the current node and spatial neighbor nodes in the initial feature distribution after linear transformation (under the k-th attention mechanism), and combined with the barrier information of the current node and spatial neighbor nodes in the barrier graph structure, a spatial attention coefficient is constructed between them, which serves as the spatial attention coefficient between the current node and spatial neighbor nodes (under the k-th attention mechanism); as follows:

[0284]

[0285] In the formula, This represents the spatial attention coefficient (combining barrier information) between the current node i and its spatial neighbor node j under the k-th attention mechanism. Let be the attention vector under the k-th attention mechanism, and be the learnable attention vector; || denotes the vector concatenation operation; σ is the activation function, preferably LeakyReLU. It should be noted that in the actual calculation, -∞ can also be replaced by a negative number with a large absolute value.

[0286] C. Normalize the spatial attention coefficients (under the k-th attention mechanism and considering the combined blocking information) corresponding to each spatial neighbor node as follows:

[0287]

[0288] In the formula, Under the k-th attention mechanism, the normalized spatial attention coefficients (integrating barrier information) between the current node i and its spatial neighbor node j; l∈{l|A spatial [i,l]=1} represents each spatial neighbor node of node i. Since exp(-∞)=0, the weight of the blocking edge is forced to 0, which is equivalent to no information being passed between the current node and the spatial neighbor nodes with blocking relationship.

[0289] D. Based on the normalized spatial attention coefficients of spatial neighbor nodes (under the k-th attention mechanism and incorporating barrier information), the initial feature vectors of each neighbor node after linear transformation (under the k-th attention mechanism) are weighted, summed, and activated to obtain the first intermediate feature vector (under the k-th attention mechanism) corresponding to the current node; as follows:

[0290]

[0291] In the formula, σ is the activation function, preferably the ELU activation function.

[0292] E. Concatenate the first intermediate feature vectors of the current node under each attention mechanism to obtain the first intermediate feature vector corresponding to the current node, as follows:

[0293]

[0294] In the formula, Wherein, K1 represents the total number of attention mechanisms in the spatial graph attention layer.

[0295] Perform steps A through E above on each node in the spatial adjacency graph to obtain the first intermediate feature vector corresponding to each node.

[0296] 3) Merge the first intermediate feature vectors corresponding to each node to obtain the first intermediate feature distribution of the sample data under the comprehensive barrier information.

[0297] Wherein, the first intermediate feature distribution satisfies

[0298] Similarly, during model training, the following steps are performed based on the semantic graph attention layer:

[0299] 1) Obtain the first intermediate feature distribution after linear transformation under different attention mechanisms;

[0300] Under different attention mechanisms, a linear transformation is performed on the first intermediate feature distribution corresponding to the sample data to obtain the first intermediate feature distribution after linear transformation with each attention mechanism;

[0301] The first intermediate feature distribution after linear transformation corresponds to the attention mechanism; that is, under different attention mechanisms, the first intermediate feature distribution after linear transformation corresponding to the attention mechanism is obtained respectively.

[0302] Specifically, under the k-th attention mechanism, a linear transformation is performed on the first intermediate feature distribution of the input, as follows:

[0303]

[0304] In the formula, Z (1)，k Let be the first intermediate feature distribution after linear transformation under the k-th attention mechanism, satisfying Where n is the total number of nodes in the current semantic similarity graph, which is the same as the total number of land parcels in the image; d head2 The feature dimension of the node in the semantic graph attention layer; Let be the weight matrix in the semantic graph attention layer under the k-th attention mechanism, and let be a trainable weight matrix that satisfies: Wherein K1×d head1 The first intermediate feature distribution Feature dimensions of each node

[0305] 2) Based on the feature vectors of the node and its corresponding semantic neighbor nodes in the first intermediate feature distribution after linear transformation, and combined with the blocking information corresponding to the node, construct the semantic attention coefficient between the two, so as to enhance the attention of the first feature vector of the node after linear transformation based on the semantic attention coefficient, and obtain the second intermediate feature distribution under the comprehensive blocking information.

[0306] Specifically, for each node in the semantic similarity graph, a second intermediate feature vector is constructed for each node based on the comprehensive blocking information, including:

[0307] A. Obtain all semantic neighbor nodes of the current node in the semantic similarity graph;

[0308] B. For each semantic neighbor node, based on the feature vectors of the current node and semantic neighbor nodes in the first intermediate feature distribution after linear transformation (under the k-th attention mechanism), and combined with the blocking information of the current node and semantic neighbor nodes in the blocking graph structure, construct the semantic attention coefficient between them, which serves as the semantic attention coefficient between the current node and the semantic neighbor node (under the k-th attention mechanism); as follows:

[0309]

[0310] In the formula, Under the k-th attention mechanism, the semantic attention coefficient (combining blocking information) between the current node i and its semantic neighbor node j; and These are the feature vectors of the current node i and its semantic neighbor node j in the first intermediate feature distribution after linear transformation (under the k-th attention mechanism); Let be the attention vector under the k-th attention mechanism, and be the learnable attention vector; || denotes the vector concatenation operation; σ is the activation function, preferably LeakyReLU. It should be noted that in the actual calculation, -∞ can also be replaced by a negative number with a large absolute value.

[0311] C. Normalize the semantic attention coefficients (under the k-th attention mechanism and considering the combined blocking information) corresponding to each semantic neighbor node as follows:

[0312]

[0313] In the formula, Under the k-th attention mechanism, the normalized semantic attention coefficients (integrating barrier information) between the current node i and its semantic neighbor node j; l∈{l|A semantic [i, l] = 1} represents each semantic neighbor of node i. Since exp(-∞) = 0, the weight of the blocking edge is forced to 0, which is equivalent to no information being passed between the current node and the semantic neighbor nodes with blocking relationship.

[0314] D. Based on the normalized semantic attention coefficients of semantic neighbor nodes (under the k-th attention mechanism and incorporating barrier information), the first intermediate feature vector (under the k-th attention mechanism) after linear transformation of each semantic neighbor node is weighted, summed, and activated to obtain the second intermediate feature vector (under the k-th attention mechanism and incorporating barrier information) corresponding to the current node; as follows:

[0315]

[0316] Where σ is the activation function, preferably the ELU activation function.

[0317] E. Concatenate the second intermediate feature vectors (comprehensive blocking information) of the current node under each attention mechanism to obtain the second intermediate feature vector (comprehensive blocking information) corresponding to the current node, as follows:

[0318]

[0319] in, Wherein, K2 represents the total number of attention mechanisms in the semantic graph attention layer.

[0320] The above steps are performed on each node in the semantic similarity graph to obtain the second intermediate feature vector (comprehensive blocking information) corresponding to each node.

[0321] 3) Merge the second intermediate feature vectors (comprehensive barrier information) corresponding to each node to obtain the second intermediate feature distribution of the sample data under the comprehensive barrier information.

[0322] in, satisfy:

[0323] To comprehensively utilize contextual information from spatial proximity graphs and semantic similarity graphs, while also considering contextual information from barrier graph structures, a feature fusion layer is used to fuse the outputs of the two graph neural network layers corresponding to the same node, obtaining the fused feature distribution of the sample data, as follows:

[0324]

[0325] in, d fused This represents the total feature dimension after concatenation.

[0326] Based on the fusion feature distribution of nodes, combined with the functional area type information corresponding to the nodes, the output layer performs functional area classification prediction on each node to obtain the functional area prediction result for each node. Specifically, this step includes:

[0327] (i) Linear transformation: High-dimensional features in the fused feature distribution are mapped to a dimension equal to the number of types C through a fully connected layer.

[0328]

[0329] in, It is a trainable weight matrix. It is the original classification score matrix of all nodes.

[0330] (ii) Probability Normalization: Z is normalized using the Softmax function to obtain the class probability vector of each node.

[0331] (iii) Prediction: Take the functional area type corresponding to the maximum value in the category probability vector of each node as its final predicted label.

[0332]

[0333] The method provided in this application constructs a barrier graph structure of geographical barrier elements to obtain the geographical barrier information of images. This geographical barrier information is then used to impose geographical barrier constraints on spatial proximity information and semantically similar information. As a result, geographical barrier information can be effectively integrated into the topological relationships of each graph structure. In other words, the geographical barrier information in the barrier graph structure is integrated into the core message passing process of the graph neural network, thereby shielding it during information aggregation. This allows the functional area identification model to fully consider geographical barrier constraints such as "functional areas are generally divided by urban main roads, main rivers, etc." during training, thereby further improving the accuracy of the functional area identification model.

[0334] To address the aforementioned technical problems in the prior art, the present invention also provides a terminal for identifying functional areas in remote sensing images to obtain information on the distribution of functional areas.

[0335] Please see Figure 4 The diagram shows the structure of the terminal; as shown below. Figure 4 As shown, the terminal 40 includes a processor 41 and a memory 42 connected to each other; the memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory, so that when the terminal is executed, it can realize all the steps of the urban land use functional area identification method based on remote sensing images as described above.

[0336] The specific steps of the urban land use functional zone identification method based on remote sensing imagery and the beneficial effects obtained by applying the terminal provided in this embodiment are the same as those in the above embodiments; they will not be repeated here.

[0337] It should be noted that the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Similarly, the processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0338] To address the aforementioned technical problems in the prior art, embodiments of the present invention also provide a computer storage medium storing a computer program that, when executed by a processor, can implement all the steps in the urban land use functional zone identification method based on remote sensing imagery as described above.

[0339] It should be noted that storage media include various media that can store program code, such as ROM, RAM, magnetic disks, or optical disks.

[0340] In summary, the urban land use functional zone identification method, terminal, and computer storage medium based on remote sensing imagery provided by this invention innovatively propose a seamless, automated functional zone identification process driven entirely by remote sensing imagery. This process begins with imagery, establishing an analytical foundation by generating land parcel patches. Then, it formally and objectively expresses the correlation knowledge based on spatial proximity and image feature similarity through the construction of a multi-level graph structure. Finally, it completes intelligent reasoning and classification through a hierarchical graph neural network, ultimately outputting the functional zone identification result. This forms a unified, vision-based solution that not only eliminates reliance on multi-source auxiliary data such as POIs and street views, and reduces dependence on massive amounts of labeled sample data, but also integrates information about geographical barriers, greatly improving the convenience, accuracy, and applicability of the functional zone identification method.

Claims

1. A method for identifying urban land use functional zones based on remote sensing imagery, characterized in that, include: Based on the preset image feature factors, the image feature vectors of each patch in the remote sensing image are extracted to construct the initial feature set corresponding to the image. By performing spatial proximity analysis and image feature similarity analysis on each locality patch, spatial proximity map and semantic similarity map corresponding to the image are constructed respectively; Based on the initial feature set, spatial proximity map, semantic similarity map and functional area information of the image, sample data corresponding to the image is constructed. The preset functional area recognition model is trained based on sample data, and the trained functional area recognition model is used to perform the functional area recognition process of the image to be recognized and obtain the corresponding functional area recognition results. The functional area identification model is a hierarchical graph neural network model constructed based on the spatial proximity graph and the semantic similarity graph.

2. The method for identifying urban land use functional zones according to claim 1, characterized in that, The method for obtaining the spatial proximity map includes: Extract the center point of each patch and set it as a node in the spatial proximity map; By calculating the degree of adjacency between the spatial locations of each node, the spatial neighbor nodes corresponding to each node are obtained; Construct the edges between a node and its corresponding spatial neighbor nodes, and set them as edges in the spatial proximity graph.

3. The method for identifying urban land use functional zones according to claim 1, characterized in that, The methods for obtaining the semantic similarity map include: Set the center point of each patch as a node in the semantic similarity graph; Set the image feature vector of the land parcel as the image feature vector of the node corresponding to the land parcel; By calculating the similarity between the image feature vectors of each node, the semantic neighbor nodes corresponding to each node are obtained; Construct the edges between a node and its corresponding semantic neighbor nodes, and set them as edges in the semantic similarity graph.

4. The method for identifying urban land use functional zones according to claim 3, characterized in that, For a single node, the methods for obtaining the semantic neighbor nodes corresponding to that node include: Calculate the similarity between the current node and all other nodes on the image feature vector to obtain the semantic similarity between the two nodes; Sort the semantic similarity between the current node and other nodes according to their numerical values; Using k-nearest neighbor selection, the k nodes with the highest scores are selected as the semantic neighbors of the current node.

5. The method for identifying urban land use functional zones according to claim 3, characterized in that, The implementation methods for constructing the edges between the node and its corresponding semantic neighbor nodes include: Construct a set of semantic neighbors for the current node by representing all the semantic neighbor nodes corresponding to the current node. For each other node in the semantic neighbor set, an undirected edge is established between the current node and other nodes, and the edge is added to the edge set; Symmetric processing is performed on each edge in the edge set, including: Perform union symmetry on two overlapping edges, or perform intersection symmetry on two overlapping edges.

6. The method for identifying urban land use functional zones according to claim 1, characterized in that, The functional area identification model includes: The input layer is used to input the initial feature set, the spatial proximity map, and the semantic similarity map of the image. The first graph neural network layer, which is connected to the input layer, is used to perform feature enhancement processing on the initial feature set based on the spatial proximity graph to obtain a first intermediate feature distribution that incorporates spatial context information. The second graph neural network layer, which connects the input layer and the first graph neural network layer, is used to perform feature enhancement processing on the first intermediate feature distribution based on the semantic similarity graph to obtain a second intermediate feature distribution that integrates the global semantic context. A feature fusion layer, which connects the first graph neural network layer and the second graph neural network layer, is used to fuse the features of the first intermediate feature distribution and the second intermediate feature distribution to obtain a fused feature distribution. The output layer, which is connected to the feature fusion layer, is used to perform classification prediction on the local land parcels in the sample data based on the fused feature distribution, and obtain the identification prediction results of the functional area.

7. The method for identifying urban land use functional zones according to claim 6, characterized in that, The first graph neural network layer uses a spatial graph convolutional layer, and the second graph neural network layer uses a semantic graph convolutional layer. Based on the spatial graph convolutional layer, the following is performed: Extract the spatial adjacency matrix corresponding to the spatial proximity graph, and calculate the degree matrix of the spatial adjacency matrix; Based on the degree matrix, a symmetric normalized graph convolution is performed on the initial feature set to obtain the first intermediate feature distribution; And, based on the semantic graph convolutional layer, the following is performed: Extract the semantic adjacency matrix corresponding to the semantic similarity graph; calculate the degree matrix of the semantic adjacency matrix; based on the degree matrix, perform symmetric normalized graph convolution on the first intermediate feature distribution to obtain the second intermediate feature distribution.

8. The method for identifying urban land use functional zones according to claim 6, characterized in that, The first graph neural network layer uses a spatial graph attention layer, and the second graph neural network layer uses a semantic graph attention layer. Based on the spatial graph attention layer, the following is performed: Obtain the initial feature distribution of the initial feature set after linear transformation under different attention mechanisms; Based on the feature vectors of a node and its corresponding spatial neighbor nodes in the initial feature distribution after linear transformation, a spatial attention coefficient is constructed between the two to enhance the attention of the node's initial feature vector after linear transformation based on the spatial attention coefficient, thereby obtaining the first intermediate feature distribution. Based on the semantic graph attention layer, the following is performed: Obtain the first intermediate feature distribution after linear transformation under different attention mechanisms; Based on the feature vectors of a node and its corresponding semantic neighbor nodes in the first intermediate feature distribution after linear transformation, a semantic attention coefficient is constructed between them. This coefficient is then used to enhance the attention of the node's first intermediate feature vector after linear transformation, thereby obtaining the second intermediate feature distribution.

9. The method for identifying urban land use functional zones according to claim 1, characterized in that, The functional area recognition process of the image to be recognized includes: Based on the preset image feature factors, extract the image feature vectors of each patch in the image to be identified in order to obtain the initial feature set corresponding to the image to be identified. By performing spatial proximity analysis and image feature similarity analysis on each locality's map patch, spatial proximity maps and semantic similarity maps corresponding to the images to be identified are constructed, respectively. Based on the initial feature set, the spatial proximity map, and the semantic similarity map of the image to be identified, the trained functional area recognition model is used to identify the image to be identified.

10. A terminal, characterized in that, include: A processor and a memory, wherein the memory and the processor are communicatively connected; The memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory, so that the terminal performs the urban land use functional zone identification method as described in any one of claims 1 to 9.

Images