Urban functional area identification method and device, electronic equipment and storage medium

Abstract

The invention relates to the technical field of geographic information, in particular to an urban functional area identification method and device, electronic equipment and a storage medium. The method comprises the following steps: acquiring multi-source data, and constructing graph structure data based on the multi-source data; and inputting the graph structure data into a preset neural network model, and outputting to obtain an urban functional area identification result, the preset neural network model being obtained by training a graph convolutional neural network model based on a spatial domain in combination with an attention mechanism fusing attribute representation and spatial connection. Therefore, city space interaction is simulated by constructing the graph structure data, and the attention mechanism jointly driven by the attributes and the space connection is utilized to identify the types of the city function areas, so that the problem of inaccurate identification of the city function areas in the prior art is solved, and accurate classification of the city function areas is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geographic information, in particular to a city function area identification method and device, an electronic device and a storage medium. BACKGROUND

[0002] City function area is a basic unit of city management and planning. The classification of city function area not only has a very important significance for city management and planning, but also provides convenience for human life, work and traffic. However, the high complexity of the city system brings great challenges to the classification of city function area. On the one hand, the spatial structure of city function area is not only determined by government decisions, but also affected by many factors such as people's daily life and market economy; on the other hand, in recent years, the process of urbanization has developed rapidly, and the spatial structure of city function area has become increasingly complex. Therefore, the research on the classification method of city function area has become a hot topic of current research.

[0003] In related technologies, city function area sample data is composed of multiple city unit sample data. The regional characteristics of city unit sample data are input into a neural network model, and the city unit sample data is divided into several types by using unsupervised learning, thereby realizing the identification of city function area.

[0004] However, the method for identifying city function area is still not accurate and needs to be solved urgently. SUMMARY

[0005] The present application provides a city function area identification method, device, electronic device and storage medium to solve the problem of inaccurate identification of city function area in the prior art, and realizes the accurate classification of city function area.

[0006] To achieve the above-mentioned purpose, the first aspect of the present application provides a city function area identification method, comprising the following steps:

[0007] Obtaining multi-source data, and constructing graph structure data based on the multi-source data;

[0008] Inputting the graph structure data into a preset neural network model to output city function area identification results, wherein the preset neural network model is trained by combining a graph convolutional neural network model based on a spatial domain with an attention mechanism that integrates attribute representation and spatial connection.

[0009] According to an embodiment of the present application, the graph structure data is constructed based on the multi-source data, comprising:

[0010] Dividing the road network data in the multi-source data to obtain a plurality of block units, and labeling the function area types of the plurality of block units by using a multi-source information visual interpretation strategy to obtain an initial function area distribution;

[0011] Preprocess the mobile phone signaling data in the multi-source data, and correspond the preprocessed mobile phone signaling data of each base station with the plurality of block units to obtain a spatial activity trajectory of a user;

[0012] Calculate an attribute representation of each block unit in the plurality of block units, wherein the attribute representation comprises at least one of a spectral feature, a texture feature, a geometric feature, and an economic semantic feature;

[0013] According to the initial functional area distribution, the spatial activity trajectory, and the attribute representation, construct the graph structure data.

[0014] According to an embodiment of the present application, the calculation of the attribute representation of each block unit in the plurality of block units comprises:

[0015] Obtain a remote sensing image, and extract the average value and the standard deviation of each band from the remote sensing image, and obtain the spectral feature according to the average value and the standard deviation of each band;

[0016] Obtain a gray level co-occurrence matrix of the remote sensing image, and obtain the texture feature based on the gray level co-occurrence matrix;

[0017] Obtain building contour data, and extract the geometric feature from the building contour data;

[0018] Construct a POI (Point Of Interest) corpus using a preset algorithm, input the POI corpus into a preset model to obtain semantic embedding of a POI category, and aggregate the semantic embedding of the POI category to obtain the economic semantic feature.

[0019] According to an embodiment of the present application, before the graph structure data is input into the preset neural network model, it further comprises:

[0020] Obtain a graph structure public data set, and divide the graph structure public data set into a training set and a validation set according to a preset ratio;

[0021] Based on a preset early stopping strategy and a preset number of iterations, train an initial neural network model using the training set, and verify the initial neural network model using the validation set until the initial neural network model meets a preset condition, end the iterative training of the initial neural network model, obtain the preset neural network model, otherwise adjust the training set and continue the iterative training.

[0022] According to one embodiment of the present application, before training the initial neural network model based on the preset early stopping strategy and the preset number of iterations using the training set, further comprising:

[0023] constructing the graph convolutional neural network model based on the spatial domain;

[0024] calculating the fusion attribute representation and spatial connection attention mechanism, and applying the fusion attribute representation and spatial connection attention mechanism to the graph convolutional neural network model based on the spatial domain to obtain the initial neural network model.

[0025] According to the urban functional area identification method proposed in the embodiments of the present application, by constructing graph structure data based on multi-source data and inputting the graph structure data into a preset neural network model, an urban functional area identification result can be output, wherein the preset neural network model is trained by a graph convolutional neural network model based on the spatial domain combined with a fusion attribute representation and spatial connection attention mechanism. Therefore, by constructing graph structure data to simulate urban space interaction and using an attribute and spatial connection driven attention mechanism to identify urban functional area types, the problem of inaccurate identification of urban functional areas in the prior art is solved, and accurate classification of urban functional areas is achieved.

[0026] To achieve the above-mentioned purpose, the second aspect of the present application proposes an urban functional area identification device, comprising:

[0027] The construction module is configured to obtain multi-source data and construct graph structure data based on the multi-source data;

[0028] The processing module is configured to input the graph structure data into a preset neural network model to output an urban functional area identification result, wherein the preset neural network model is trained by a graph convolutional neural network model based on the spatial domain combined with a fusion attribute representation and spatial connection attention mechanism.

[0029] According to one embodiment of the present application, the construction module comprises:

[0030] The first obtaining unit is configured to divide road network data in the multi-source data to obtain a plurality of block units, and label the plurality of block units with functional area types using a multi-source information visual interpretation strategy to obtain an initial functional area distribution;

[0031] The second obtaining unit is configured to preprocess mobile phone signaling data in the multi-source data, and correspond the preprocessed mobile phone signaling data of each base station with the plurality of block units to obtain spatial activity tracks of users;

[0032] a calculation unit, configured to calculate attribute representation of each of the plurality of block units, wherein the attribute representation comprises at least one of spectral features, texture features, geometric morphological features, and economic semantic features;

[0033] a construction unit, configured to construct the graph structure data according to the initial functional area distribution, the spatial activity trajectory, and the attribute representation.

[0034] According to an embodiment of the present application, the calculation unit is specifically configured to:

[0035] obtain remote sensing images, and extract the average value and the standard deviation of each band from the remote sensing images, and obtain the spectral features according to the average value and the standard deviation of each band;

[0036] obtain a gray level co-occurrence matrix of the remote sensing images, and obtain the texture features based on the gray level co-occurrence matrix;

[0037] obtain building contour data, and extract the geometric morphological features from the building contour data;

[0038] construct a POI corpus by using a preset algorithm, input the POI corpus into a preset model to obtain semantic embedding of POI categories, and aggregate the semantic embedding of the POI categories by weighting to obtain the economic semantic features.

[0039] According to an embodiment of the present application, before the graph structure data is input into the preset neural network model, the processing module further comprises:

[0040] an acquisition unit, configured to acquire a graph structure public data set, and divide the graph structure public data set into a training set and a verification set according to a preset ratio;

[0041] a training unit, configured to train an initial neural network model by using the training set based on a preset early stopping strategy and a preset iteration number, and verify the initial neural network model by using the verification set, until the initial neural network model meets a preset condition, end the iterative training of the initial neural network model, obtain the preset neural network model, or otherwise continue the iterative training after adjusting the training set.

[0042] According to an embodiment of the present application, before the initial neural network model is trained by using the training set based on the preset early stopping strategy and the preset iteration number, the training unit is further configured to:

[0043] construct the graph convolutional neural network model based on the spatial domain;

[0044] The fusion attribute representation and the spatial connection attention mechanism are connected, and the fusion attribute representation and the spatial connection attention mechanism are applied to the spatial domain-based graph convolutional neural network model to obtain the initial neural network model.

[0045] According to the urban function area identification device provided by the embodiment of the present application, the graph structure data is constructed based on multi-source data, and the graph structure data is input into the preset neural network model, so that the urban function area identification result can be output. The preset neural network model is trained by combining a spatial domain-based graph convolutional neural network model with a fusion attribute representation and a spatial connection attention mechanism. Therefore, by constructing the graph structure data to simulate the urban space interaction and using the attention mechanism driven by the attributes and the spatial connection, the urban function area type is identified, and the problem of inaccurate identification of the urban function area in the prior art is solved, and the accurate classification of the urban function area is realized.

[0046] To achieve the above object, the third aspect of the present application provides an electronic device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the urban function area identification method as described in the above embodiments.

[0047] To achieve the above object, the fourth aspect of the present application provides a computer readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to implement the urban function area identification method as described in the above embodiments.

[0048] The additional aspects and advantages of the present application will be partially given in the following description, partially will become obvious from the following description, or will be understood by the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0049] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which:

[0050] Figure 1 A flowchart of an urban function area identification method according to an embodiment of the present application;

[0051] Figure 2 A flowchart of another urban function area identification method according to an embodiment of the present application;

[0052] Figure 3 A classification result schematic diagram of an attribute feature selection comparison experiment according to an embodiment of the present application;

[0053] Figure 4 A content similarity-based attention mechanism calculation process schematic diagram according to an embodiment of the present application;

[0054] Figure 5 Fig. 6 is a diagram illustrating the impact of structural details according to an embodiment of the present application on attention coefficients;

[0055] Figure 6 Fig. 7 is a diagram illustrating spatial activity trajectory weight calculation according to an embodiment of the present application;

[0056] Figure 7 Fig. 8 is a diagram illustrating classification results of comparative experiments of attention strategies according to an embodiment of the present application;

[0057] Figure 8 Fig. 9 is a diagram illustrating classification results of comparative experiments with existing intelligent methods according to an embodiment of the present application;

[0058] Figure 9 Fig. 10 is a block diagram of a city functional area identification apparatus according to an embodiment of the present application;

[0059] Figure 10 Fig. 11 is a structural diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0060] Embodiments of the present application are described in detail below with reference to the accompanying drawings. Examples of the embodiments are shown in the drawings, in which the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the drawings are exemplary and are intended to explain the present application, and should not be understood as limiting the present application.

[0061] A city functional area identification method, apparatus, electronic device and storage medium according to embodiments of the present application are described below with reference to the accompanying drawings.

[0062] Figure 1 Fig. 1 is a flowchart of a city functional area identification method according to an embodiment of the present application.

[0063] Before introducing the city functional area identification method according to embodiments of the present application, the related technical background is introduced.

[0064] The formation of urban functional zones is the result of the joint action of complex geographical environment and human production and living activities. Different functional zones show significant differences in geographical information, socio-economic information and human activities in multiple dimensions. Therefore, scholars now identify urban functional zones from the perspective of classification. In existing research, scholars use remote sensing image data, urban building vector data and urban perception data to study urban functional zones. Among them, remote sensing image data can extract spectral, texture and other physical information to finely depict the internal spatial pattern of urban functional zones; building vector data can map the type characteristics of functional zones by analyzing the shape and geometric structure of buildings. However, both types of data have certain limitations in describing the internal socio-economic information of functional zones. In this regard, some scholars began to explore the economic information hidden in urban perception data to assist in functional zone classification. However, these methods are limited to the internal characteristics of functional zones and cannot fully reflect the characteristics of human activities in the region, and there is still a lack of expression of spatial interaction within the city. With the rapid development of travel big data technology, the spatial activity information it contains provides strong support for revealing the spatial interaction relationship of urban areas. A large number of experiments have shown that the classification method of urban functional zones combined with travel big data and geographical information data can more accurately reflect the attribute representation and spatial activity characteristics of urban areas, and thus improve the accuracy of functional zone classification.

[0065] In related technologies, the methods for studying urban functional zone classification can be roughly divided into three categories, namely, the urban functional zone classification method based on attribute representation, the urban functional zone classification method based on spatial activity, and the urban functional zone classification method based on graph convolutional neural network.

[0066] (1) The urban functional zone classification method based on attribute representation uses remote sensing image data, urban building vector data and urban perception data to extract spectral, texture and other physical information to finely depict the internal spatial pattern of urban functional zones, building shape and geometric structure to map the type characteristics of functional zones, and POI to describe the economic semantic characteristics of functional zones. Socio-economic information integrates multi-source data to coexist and analyze the landscape characteristics and economic activity characteristics within the functional zone, achieving a classification closer to the overall characteristics of the functional zone.

[0067] (2) The urban functional area classification method based on spatial activities uses the spatial activity information reflecting the social dynamic functional attributes of the region contained in the travel big data to understand the various influencing factors affecting the spatio-temporal distribution of residents' travel by studying the travel characteristics of residents in the functional area, such as obtaining various travel data such as taxis, bicycles, mobile phone signaling, etc., obtaining the changes of various travel indicators at different time points in the functional area such as travel growth, outflow, peak hours, etc. through spatial statistics, sample estimation method, clustering algorithm, etc. to count the residents' activity patterns to show the dynamic activity characteristics of the functional area, thereby providing a reference for the classification of functional areas. By mining the spatial interaction characteristics between urban regions, a spatial interaction network between urban regions is constructed to further mine the flow and demand of human beings, reveal the influence of human activities on the type of functional area, and the attributes and spatial activities both reflect the urban function. The former provides the static attributes of the region, while the latter provides the dynamic characteristics of the region and the spatial interaction between regions. Considering the complementary characteristics of both, many studies combine the use of user travel big data and geographic spatial data to perform the task of urban functional area classification, such as combining POI and trajectory data to use the document topic generation model LDA (Latent Dirichlet Allocation, also known as a three-layer Bayesian probability model) or Word2Vec to perform functional area classification, and using remote sensing images and mobile phone trajectory data to extract and integrate landscape indicators and human activity indicators to identify urban functional areas through hierarchical clustering.

[0068] (3) The urban functional area classification method based on graph convolutional neural network represents each urban area as a node in the graph structure data, establishes the graph structure data by establishing the spatial correlation between urban areas as the edges between nodes, and uses the graph convolution network model to analyze the characteristics of urban areas and between urban areas. In existing research, there are two methods to establish graph structure data: one is to establish a graph structure for objects (buildings, etc.) within the functional area, and to classify the functional area according to the differences in the graph structure of different functional areas; the second is to establish a graph structure between functional areas according to the spatial interaction of functional areas. Most existing researches adopt the first method, that is, considering different spatial structure patterns within the functional area, focusing on the spatial relationship between internal elements, for example, according to the first law of geography, topologically adjacent areas are connected. However, studies have shown that the flow connection of human spatial activities is more accurate than the flow connection of topologically adjacent areas, and the city is a complex system containing people and nature, the interaction between humans and the city has an important influence on the determination of urban form, selecting the trajectory of human spatial activity as the connection between research units can obtain additional spatial context and the strength and pattern of the connection between different areas through the connection information of the spatial activity trajectory. However, existing researches have not effectively integrated the attribute representation of the research unit and the spatial activity connection between the research units. In addition, according to the different information acquisition methods, the graph convolution network is divided into spectral domain and spatial domain based models, and in terms of comprehensive efficiency, universality and flexibility, it is found that the spatial domain based model is superior to the spectral domain based model.

[0069] Therefore, the embodiment of the present application proposes an urban functional area identification method, which simulates urban spatial interaction by constructing graph structure data, and identifies the type of urban functional area by using an attention mechanism driven by attributes and spatial connections, solving the problem of inaccurate identification of urban functional areas in the prior art, and realizing accurate classification of urban functional areas.

[0070] For example, as shown in Figure 1 The urban functional area identification method includes the following steps:

[0071] In step S101, multi-source data is obtained, and graph structure data is constructed based on the multi-source data.

[0072] The multi-source data includes road network data (such as road centerline or road boundary), mobile signaling data, remote sensing image data, building outline data and POI data.

[0073] Specifically, for road network data, it can be obtained by the National Geographic Information Center and map service providers (such as Baidu Map); mobile signaling data includes signal strength, location change and other information of mobile phones, which can be collected by telecom operators; remote sensing image data can be obtained by purchasing or subscribing to satellite and airline image services, or downloading free remote sensing image data provided by open source data platforms; building outline data can be obtained from city planning departments or building information models; POI data can be obtained from map service providers and open data platforms. After obtaining multi-source data, the data can be integrated and constructed into graph structure data, which consists of nodes and connecting edges.

[0074] For ease of understanding, how to construct graph structure data based on multi-source data is described in detail below.

[0075] As a possible implementation manner, in some embodiments, constructing graph structure data based on multi-source data comprises: dividing road network data in the multi-source data to obtain a plurality of block units, and labeling the plurality of block units with function area types by using a multi-source information visual interpretation strategy to obtain an initial function area distribution; preprocessing mobile signaling data in the multi-source data, and corresponding the preprocessed mobile signaling data of each base station with the plurality of block units to obtain spatial activity trajectories of users; calculating attribute representations of each block unit in the plurality of block units, wherein the attribute representations include at least one of spectral features, texture features, geometric morphological features and economic semantic features; and constructing graph structure data according to the initial function area distribution, the spatial activity trajectories and the attribute representations.

[0076] Specifically, as Figure 2As shown, the embodiments of the present application can adopt technologies such as but not limited to ARCGIS software, based on network analysis and spatial overlay, etc. to divide the road network data in the multi-source data into block units to obtain a plurality of block units, and based on the multi-source information visual interpretation strategy, to judge the function area type (such as commercial area, residential area, industrial area, etc.) of the plurality of block units after division. For the block unit with mixed functions, the function type with the largest area ratio can be selected to represent the function type of the block unit, or the relative importance of the function type is comprehensively considered, so as to obtain the initial function area distribution; the mobile phone signaling data is preprocessed and associated with the block unit to determine the block unit corresponding to the mobile phone signaling data of each base station, and the user's stopping points are identified through data cleaning strategies (such as denoising, filtering and interpolation, etc.), and these stopping points are connected to form the user's spatial activity trajectory, while eliminating the invalid trajectories (such as frequent switching of users under the same base station) between the mobile phone signal base stations in the block unit and the outliers (which may be caused by equipment failure or signal interference) exceeding the speed threshold; the attribute representation of each block unit in the plurality of block units is calculated, including urban landscape features and economic semantic features, wherein the urban landscape features include spectral features, texture features and geometric morphological features, based on the above initial function area distribution, spatial activity trajectory and attribute representation, a graph structure data G=(V, E, N) is constructed, wherein V is a node set, E is an edge set, and N is an edge weight set. In the node set V, each node represents a block unit, the attribute representation of the block unit is the feature vector of the node, and the set of spatial activity trajectories of the user is the edge set E. In the edge weight set N, each edge connecting two nodes has an edge weight determined by the number of times of the spatial activity trajectory represented by the edge.

[0077] The following will explain in detail how to calculate the attribute representation of each block unit in the plurality of block units.

[0078] Optionally, in some embodiments, calculating the attribute representation of each block unit in the plurality of block units comprises: obtaining a remote sensing image, and extracting the average value and standard deviation of each band from the remote sensing image, and obtaining the spectral features according to the average value and standard deviation of each band; obtaining a gray level co-occurrence matrix of the remote sensing image, and obtaining the texture features based on the gray level co-occurrence matrix; obtaining building contour data, and extracting the geometric morphological features from the building contour data; constructing a POI corpus using a preset algorithm, inputting the POI corpus into a preset model to obtain semantic embeddings of POI categories, and weighting and aggregating the semantic embeddings of POI categories to obtain economic semantic features.

[0079] Specifically, the spectral features can be extracted from the average value and standard deviation of each band of the remote sensing image, which reflect the brightness distribution and variation range of the block unit in different spectral bands; based on the texture analysis method such as the gray level co-occurrence matrix of the remote sensing image, the texture features of the block unit can be extracted, which describe the spatial distribution pattern and complexity of the land cover type; the geometric morphological features of the block unit are extracted from the building contour data, such as boundary length, shape index, compactness, etc., which reflect the spatial form and layout characteristics of the block unit; the POI corpus is constructed by using a preset algorithm (such as random walk algorithm) in the POI data, the moving path between different POIs is simulated, the POI corpus is input into a preset model (such as Skip-Gram model), and the semantic embedding representation of the POI category is learned, and the economic semantic feature vector of the block unit is obtained by weighted aggregation according to the number and distribution of POI categories in the block unit.

[0080] In step S102, the graph structure data is input into a preset neural network model, and a city functional area recognition result is output. The preset neural network model is trained based on a spatial domain graph convolutional neural network model combined with an attention mechanism that fuses attribute representation and spatial connection.

[0081] That is, as shown in Figure 2 After obtaining the graph structure data, the graph structure data is input into a preset neural network model (such as AEGCNN (Autoencoder Graph Convolutional Neural Network, automatic encoder-graph convolutional neural network)), and a city functional area recognition result can be output. The embodiment of the present application solves the problem of city functional area recognition by using a spatial domain graph convolutional neural network model with an attention mechanism that is jointly driven by attribute representation and spatial connection. In order to better understand the spatial domain graph convolutional neural network model and the attention mechanism, the following will introduce the related theoretical knowledge of graph convolutional neural network and attention mechanism.

[0082] As a natural extension of convolutional neural networks to non-Euclidean data, graph convolutional neural networks can convert practical problems into connection and message propagation problems between nodes, model the interaction between graph nodes, and achieve deep processing of graph structure data. Not only can it extract feature information of graph structure data, but also can capture its structure information, and then complete point classification, graph classification, edge prediction and graph embedding representation and other tasks. Graph convolutional neural networks can be divided into two categories based on spectral domain and spatial domain. The embodiment of the present application uses a graph convolutional neural network based on spatial domain. Compared with the spectral domain model, the spatial domain model directly performs convolution operation in the graph structure through information propagation. Its graph convolution calculation can be performed in a batch of node subgraphs, rather than in the entire graph, and thus is more suitable for processing large graphs. First, a certain number of neighbor nodes are sampled for each center node according to a certain rule. In each iteration of graph neural network training, the neighbor nodes can remain unchanged or be resampled. Then, an aggregation function is used to calculate the aggregation information of the neighbor nodes and the center node to obtain the embedding representation of the center node and update the information of the center node.

[0083] Attention mechanism in machine learning is a technology that simulates human attention behavior. Its core goal is to select more important information for the current task from a large amount of information, and by giving different weights to different parts of the information, the model can focus on important information, thereby improving processing efficiency and accuracy. The advantage of attention mechanism is that it does not treat input information equally, but automatically selects and weights different parts of the input according to task requirements. The weighting process usually involves calculating a set of attention weights, which represent the degree of attention of the model to different positions in the input information sequence. There are many types of attention mechanisms, including global attention, local attention, self-attention, and multi-head attention. The multi-head attention mechanism used in the embodiment of the present application allows the model to simultaneously learn multiple sets of different queries, keys and values in order to perform attention calculation in different subspaces. In processing sequence data, attention mechanism can obtain these weights by calculating the similarity between query vector (Query) and key vector (Key), so as to determine which parts of the input sequence should be focused on at each position of the output sequence. Attention weights are obtained by comparing query vector and key vector through a score function, and then normalized to obtain the original attention score, which is then multiplied by the value vector of the input sequence itself to obtain the weighted sum as the output of the attention mechanism:

[0084]

[0085] where d k is the dimension of the key vector, q is the query vector, k is the key vector, and v is the value vector.

[0086] Under the action of the attention mechanism, the graph convolutional network can dynamically calculate the corresponding attention weight according to the features of the current node and the neighbor nodes, so that the model can flexibly adjust the allocation of attention between different nodes and time steps, and thus better capture the local and global information in the graph structure:

[0087]

[0088] wherein H is the output node feature matrix, representing the updated features of the nodes after the graph convolutional layer; σ is an activation function, usually a ReLU function or a Leaky ReLU function; N is the number of nodes in the graph structure data; W is a training parameter matrix; a ij is the attention weight, representing the attention weight coefficient between node i and node j; A is the adjacency matrix of the graph; X is the feature vector matrix of the graph nodes.

[0089] As an example, the following describes in detail how to obtain the preset neural network model.

[0090] Optionally, in some embodiments, before inputting the graph structure data into the preset neural network model, the method further comprises: obtaining a graph structure public data set, and dividing the graph structure public data set into a training set and a validation set according to a preset ratio; based on a preset early stopping strategy and a preset number of iterations, training the initial neural network model using the training set, and verifying the initial neural network model using the validation set, until the initial neural network model meets a preset condition, ending the iterative training of the initial neural network model, obtaining the preset neural network model, otherwise adjusting the training set and continuing the iterative training.

[0091] Specifically, in a pytorch deep learning environment based on a python programming language, the graph structure public data set is input into the initial neural network model for training and verification, and the preset neural network model can be obtained. The aforementioned graph structure data is input into the preset neural network model, and the city functional area recognition result can be obtained. The graph structure public data set is divided into a training set and a validation set, wherein the training set: test set = 3:1, a total of 5 repeated experiments are performed, the number of training rounds in each experiment is set to 500 (i.e. the preset number of iterations), the preset early stopping strategy is used, the training is stopped when the accuracy rate is not rising in the specified number of rounds, the model classification accuracy rate is calculated according to the obtained city functional area recognition result, and the Mirco-F1 coefficient and the Kappa coefficient are used as the evaluation standard of the model classification accuracy rate, wherein the Mirco-F1 coefficient is used to measure the comprehensive situation of the precision (Precision) and the recall (Recall) of the classifier in the multi-class classification task, and the calculation formula is:

[0092]

[0093] where TP i is the number of correct classifications, FP i is the number of other classes judged as the i-th class, FN i is the number of the i-th class judged as other classes.

[0094] The Kappa coefficient is a statistical measure of agreement, with a range of [-1, 1], suitable for scoring systems and classification problems, used to measure the consistency between model output results and actual results. The larger the Kappa coefficient, the higher the consistency between model output results and actual results:

[0095]

[0096] where a1, a2,... a n is the number of each class of real label, b1, b2,... b n is the number of each class of predicted label, n is the total number of samples, and p0 is the sum of the number of correct classifications of each class divided by the total number.

[0097] When the value of the Kappa coefficient is negative, it means that the model output result is lower than the random classification result. When the value of the Kappa coefficient is positive, the larger the value, the stronger the consistency between the model output result and the actual result. When the value of the Kappa coefficient is 0, it means that the model output result is coincidentally consistent with the actual result.

[0098] The classification accuracy and classification result confusion matrix of five experiments are analyzed to evaluate on objective training data, and the good accuracy and robustness of the graph convolution model in the urban functional area classification problem are determined. As Figure 3 shown, to verify the rationality of the selection of street unit node features, comparative experiments are conducted, setting three groups of controls: one is to use only urban landscape features (such as Figure 3 shown in (c)); two is to use only economic semantic features (such as Figure 3 shown in (d)); three is the co-occurrence analysis of urban landscape features and economic semantic features (such as Figure 3 shown in (b)). The Micro-F1 and Kappa values of the classification results of the three experiments show that when the urban landscape features and economic semantic features are fused and analyzed, both are significantly improved, showing the effectiveness of the model in classifying street units in areas with similar spatial distribution patterns and less POI data in the urban fringe.

[0099] The following describes how to obtain the initial neural network.

[0100] Optionally, in some embodiments, before training the initial neural network model based on the preset early stopping strategy and the preset number of iterations using the training set, the method further comprises: constructing a graph convolutional neural network model based on a spatial domain; calculating a fusion attention mechanism of fused attribute representation and spatial connection, and applying the fusion attention mechanism of fused attribute representation and spatial connection to the graph convolutional neural network model based on the spatial domain to obtain the initial neural network model.

[0101] Specifically, the graph convolutional neural network model based on the spatial domain constructed by the embodiment of the application comprises three graph convolutional layers, which are used to process the input graph structure data G=(A, X). Each graph convolutional layer selects a first-order neighborhood as a convolutional neighborhood, that is, each time the graph convolutional operation is performed, the neighbor nodes of each center node are selected for sampling. In the training process, the training parameters are iteratively updated through a large number of back propagation operations based on the error between the node results output by each layer of the graph convolutional neural network and the one-hot encoding of the given label, so as to continuously optimize the performance of the model. Finally, the graph convolutional model can complete the functional area classification by using the node classification task of the graph structure data; a fusion attention mechanism of fused attribute representation and spatial connection is calculated and applied to the graph convolutional neural network model based on the spatial domain to obtain the initial neural network model. The attribute representation is used to measure the content similarity between nodes, and the spatial connection comprises two dimensions of connection strength defined by the structural importance of the graph and the spatial activity.

[0102] In combination with Figures 4-6 As shown in the figure, the process of calculating the fusion attention mechanism of fused attribute representation and spatial connection is as follows:

[0103] (1) Calculate the content similarity between node v i and node v j .

[0104] A fea (Wh i , Wh j )=a T (Wh i || Wh j )

[0105] Where || represents a feature splicing operation, W is a training parameter for mapping node features to a high-dimensional space, A fea (f i , f j ) is a similarity between two attribute representations f i and f j . The higher the content similarity, the more similar the attribute representations of the two nodes, and the higher the attention weight that can be obtained.

[0106] (2) Calculate the content similarity between node v i and node v jThe structural importance between them, through A str (f i ,f j Quantify the structural importance between two nodes, at node v i The structural receptive field is calculated using a restart random walk algorithm to obtain the node weights w. ip Node v j Similarly, weighted Jacard similarity is used to evaluate structural importance:

[0107]

[0108] The structural receptive field is a subgraph (v) containing all nodes within a node's k-th order neighbors, including the node itself. i E i This paper proposes a method to extract dense node clusters around a node by constructing a structural receptive field, thereby mining structural details and integrating these details into the graph's attention mechanism to support subsequent node classification tasks. To obtain the structural importance between nodes within the structural receptive field, this embodiment employs a restarted random walk algorithm. The second-order neighborhood of a node is selected as the structural receptive field, and the restarted random walk algorithm captures structural details within this field. The restart probability of the algorithm is 0.5. When a central node is close to a densely distributed node cluster, the restarted random walk algorithm assigns higher weights to the dense cluster nodes connected to the central node, thus effectively summarizing the characteristics of the graph structure data.

[0109] (3) Calculate node v i and node v j The spatial connection strength between nodes, for each node v in the graph structure data i Its neighbor node v j (j∈N i Spatial connection strength between ) Spt ij It is through computing node v i With node v j The number of times n of human activity trajectories occur between them ij and divide by node v i The sum of the number of times its trajectory appears with all its neighboring nodes:

[0110]

[0111] Where, N i For node v i The set of all neighboring nodes; the greater the spatial connectivity, the greater the weight of the spatial activity trajectory.

[0112] (4) Normalize content similarity, structural importance, and spatial connectivity strength, and calculate the final fusion attention coefficient w.ij :

[0113]

[0114] wherein a(·), b(·) and g(·) can be functions, before combining the content similarity, the structural importance and the spatial connection strength, the weights of which are adjusted, for the convenience of operation, the application embodiments use constants a = 0.4, b = 0.3 and g = 0.3.

[0115] The fusion attention mechanism calculates the fusion of the attribute representation and the spatial connection, that is, when determining the weight of the node, not only the inherent attribute representation is considered, but also the information in the spatial activity dimension is fully considered, so that the calculation of the attention coefficient is more comprehensive and accurate.

[0116] Further, in order to verify the effectiveness and superiority of the application embodiments, comparative experiments of attention integration strategy and existing intelligent methods are carried out. As shown in Figure 7 , the comparative experimental results of the three attention strategies used by the model in the classification accuracy verify the advantages of the fusion attention strategy used in the application embodiments in the urban functional area identification task, wherein the three attention strategies are the attribute-enhanced attention strategy (as shown in Figure 7 (c)), the connection-enhanced attention strategy (as shown in Figure 3 (d)), and the fusion attribute representation and spatial connection-enhanced attention strategy used by AE GCNN (as shown in Figure 3 (b)). As shown in Figure 8 , by comparing and analyzing the classification accuracy of AE GCNN (as shown in Figure 8 (b)) and other commonly used intelligent methods, the performance advantage of AE GCNN in the urban functional area classification task can be verified, wherein the other commonly used intelligent methods include the basic graph convolutional network (GCN) (as shown in Figure 8 (c)) and the support vector machine (SVM) (as shown in Figure 8 (d)).

[0117] In summary, the application embodiments have the following advantages:

[0118] (1) Multiple sources of big data are comprehensively used to construct a graph structure data with street units as graph nodes and spatial activity trajectories as connection edges, and urban landscape features and economic semantic features are selected to completely describe the attribute representation of the research unit, so that the input features of the model are close to expert knowledge to some extent, to ensure the accuracy of urban functional area classification;

[0119] (2) A graph convolutional neural network model AE GCNN based on a spatial domain fusion attention mechanism is constructed, and an attribute and connection jointly driven attention mechanism is used for graph convolution operation to consider more dimensional information.

[0120] According to the urban function area identification method provided in the embodiments of the present application, the graph structure data is constructed based on the multi-source data, and the graph structure data is input into the preset neural network model to output the urban function area identification result, wherein the preset neural network model is trained based on the graph convolutional neural network model based on the spatial domain and the attention mechanism of the fused attribute representation and spatial connection. Therefore, the urban function area type is identified by constructing the graph structure data to simulate the urban space interaction and using the attention mechanism jointly driven by the attribute and spatial connection, and the problem of inaccurate identification of the urban function area in the prior art is solved, and the accurate classification of the urban function area is realized.

[0121] Secondly, the urban function area identification device provided in the embodiments of the present application is described with reference to the accompanying drawings.

[0122] Figure 9 is a block schematic diagram of the urban function area identification device in one embodiment of the present application.

[0123] As shown in Figure 9 , the urban function area identification device 10 comprises a construction module 100 and a processing module 200.

[0124] The construction module 100 is configured to acquire multi-source data and construct graph structure data based on the multi-source data.

[0125] The processing module 200 is configured to input the graph structure data into a preset neural network model to output an urban function area identification result, wherein the preset neural network model is trained based on a graph convolutional neural network model based on a spatial domain and an attention mechanism of a fused attribute representation and spatial connection.

[0126] Further, in some embodiments, the construction module 100 comprises:

[0127] A first obtaining unit is configured to divide road network data in the multi-source data to obtain a plurality of block units, and label the function area types of the plurality of block units by using a multi-source information visual interpretation strategy to obtain an initial function area distribution;

[0128] A second obtaining unit is configured to preprocess mobile signaling data in the multi-source data, and correspond the preprocessed mobile signaling data of each base station with the plurality of block units to obtain spatial activity tracks of users.

[0129] The computing unit is configured to calculate attribute representations of each of the plurality of block units, wherein the attribute representations comprise at least one of spectral features, texture features, geometric morphological features, and economic semantic features.

[0130] The constructing unit is configured to construct graph structure data according to the initial functional area distribution, the spatial activity trajectory, and the attribute representations.

[0131] Further, in some embodiments, the computing unit is specifically configured to:

[0132] The remote sensing image is acquired, and the average value and the standard deviation of each band are extracted from the remote sensing image, and the spectral features are obtained according to the average value and the standard deviation of each band.

[0133] The gray level co-occurrence matrix of the remote sensing image is acquired, and the texture features are obtained based on the gray level co-occurrence matrix.

[0134] The building contour data is acquired, and the geometric morphological features are extracted from the building contour data.

[0135] The POI corpus is constructed by using a preset algorithm, the semantic embedding of the POI category is obtained by inputting the POI corpus into a preset model, and the economic semantic features are obtained by weighting and aggregating the semantic embedding of the POI category.

[0136] Further, in some embodiments, before the graph structure data is input into the preset neural network model, the processing module 200 further comprises:

[0137] The acquiring unit is configured to acquire a graph structure public data set, and divide the graph structure public data set into a training set and a validation set according to a preset ratio.

[0138] The training unit is configured to train the initial neural network model by using the training set based on a preset early stopping strategy and a preset iteration number, and verify the initial neural network model by using the validation set, until the initial neural network model meets a preset condition, end the iterative training of the initial neural network model, obtain the preset neural network model, or otherwise continue the iterative training after adjusting the training set.

[0139] Further, in some embodiments, before the initial neural network model is trained by using the training set based on the preset early stopping strategy and the preset iteration number, the training unit is further configured to:

[0140] The graph convolutional neural network model based on the spatial domain is constructed.

[0141] The attention mechanism fusing the attribute representations and the spatial connections is calculated, and the attention mechanism fusing the attribute representations and the spatial connections is applied to the graph convolutional neural network model based on the spatial domain to obtain the initial neural network model.

[0142] It should be noted that the foregoing explanation and description of the embodiment of the city function area identification method also apply to the city function area identification device of this embodiment, which will not be described here.

[0143] According to the city function area identification device provided by the embodiment of the present application, the graph structure data is constructed based on multi-source data, and the graph structure data is input into the preset neural network model, so that the city function area identification result can be output, wherein the preset neural network model is trained by combining the graph convolutional neural network model based on the spatial domain with the attention mechanism of fusion attribute representation and spatial connection. Therefore, by constructing the graph structure data to simulate the city space interaction, and using the attention mechanism driven by the attribute and the spatial connection together to identify the city function area type, the problem of inaccurate identification of the city function area in the prior art is solved, and the accurate classification of the city function area is realized.

[0144] Figure 10 The structural schematic diagram of the electronic device provided by the embodiment of the present application is shown. The electronic device can include:

[0145] The memory 1001, the processor 1002 and the computer program stored in the memory 1001 and executable on the processor 1002.

[0146] The processor 1002 executes the program to realize the city function area identification method provided in the above embodiment.

[0147] Further, the electronic device further includes:

[0148] The communication interface 1003 is used for communication between the memory 1001 and the processor 1002.

[0149] The memory 1001 is used to store the computer program executable on the processor 1002.

[0150] The memory 1001 can contain a high-speed RAM (Random Access Memory, Random Access Memory) memory, and can also include a non-volatile memory, such as at least one disk memory.

[0151] If the memory 1001, the processor 1002 and the communication interface 1003 are implemented independently, the communication interface 1003, the memory 1001 and the processor 1002 can be connected with each other through a bus and complete communication between each other. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, Figure 10 Only one thick line is used in the figure to represent the bus, but it does not mean that there is only one bus or only one type of bus.

[0152] Optionally, in a specific implementation, if the memory 1001, the processor 1002 and the communication interface 1003 are integrated on a chip, the memory 1001, the processor 1002 and the communication interface 1003 can complete communication between each other through an internal interface.

[0153] The processor 1002 can be a CPU (Central Processing Unit), or an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement one or more embodiments of the present application.

[0154] The embodiments of the present application also provide a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the above urban functional area identification method.

[0155] In addition, the terms "first", "second" are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "multiple" is at least two, for example, two, three, etc., unless otherwise explicitly and specifically limited.

[0156] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are contained in at least one embodiment or example of the present application. In the specification, the illustrative description of the above terms is not necessarily directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the specification and the features of the different embodiments or examples without contradiction.

[0157] Although the embodiments of the present application have been shown and described above, it is understood that the above-described embodiments are exemplary and are not to be construed as limiting the present application, and the person skilled in the art can make changes, modifications, replacements and variations to the above-described embodiments within the scope of the present application.

Claims

1. A method for identifying urban functional zones, characterized in that, Includes the following steps: Acquire multi-source data and construct graph structure data based on the multi-source data; The graph structure data is input into a preset neural network model, and the output is the urban functional area identification result. The preset neural network model is trained by combining a spatial domain-based graph convolutional neural network model with an attention mechanism that integrates attribute representation and spatial connectivity.

2. The method according to claim 1, characterized in that, The construction of graph structure data based on the multi-source data includes: The road network data in the multi-source data is divided into multiple street blocks, and the functional area type is labeled for the multiple street blocks using a multi-source information visual interpretation strategy to obtain the initial functional area distribution. The mobile phone signaling data in the multi-source data is preprocessed, and the preprocessed mobile phone signaling data of each base station is mapped to the multiple street blocks to obtain the user's spatial activity trajectory. Calculate the attribute representation of each of the plurality of block units, wherein the attribute representation includes at least one of spectral features, texture features, geometric morphology features and economic semantic features; The graph structure data is constructed based on the initial functional area distribution, the spatial activity trajectory, and the attribute representation.

3. The method according to claim 2, characterized in that, The calculation of the attribute representation of each of the plurality of block units includes: Acquire remote sensing images, and extract the mean and standard deviation of each band from the remote sensing images, and obtain the spectral features based on the mean and standard deviation of each band; Obtain the gray-level co-occurrence matrix of the remote sensing image, and obtain the texture features based on the gray-level co-occurrence matrix; Obtain building outline data, and extract the geometric features from the building outline data; A POI corpus is constructed using a preset algorithm. The POI corpus is then input into a preset model to obtain the semantic embeddings of POI categories. The semantic embeddings of the POI categories are then weighted and aggregated to obtain the economic semantic features.

4. The method according to claim 1, characterized in that, Before inputting the graph structure data into the preset neural network model, the method further includes: Obtain a public dataset of graph structures, and divide the public dataset of graph structures into a training set and a validation set according to a preset ratio; Based on a preset early stopping strategy and a preset number of iterations, the initial neural network model is trained using the training set and validated using the validation set until the initial neural network model meets preset conditions. Then, the iterative training of the initial neural network model ends, and the preset neural network model is obtained. Otherwise, the training set is adjusted and iterative training continues.

5. The method according to claim 4, characterized in that, Before training the initial neural network model using the training set based on the preset early stopping strategy and the preset number of iterations, the method further includes: Construct the spatial domain-based graph convolutional neural network model; The attention mechanism of the fused attribute representation and spatial connectivity is calculated, and the attention mechanism of the fused attribute representation and spatial connectivity is applied to the spatial domain-based graph convolutional neural network model to obtain the initial neural network model.

6. A device for identifying urban functional zones, characterized in that, include: A construction module is used to acquire multi-source data and construct graph structure data based on the multi-source data; The processing module is used to input the graph structure data into a preset neural network model and output the urban functional area identification result. The preset neural network model is trained by combining a spatial domain-based graph convolutional neural network model with an attention mechanism that integrates attribute representation and spatial connectivity.

7. The apparatus according to claim 6, characterized in that, The building module includes: The first obtaining unit is used to divide the road network data in the multi-source data to obtain multiple block units, and to use a multi-source information visual interpretation strategy to label the multiple block units with functional area types to obtain the initial functional area distribution. The second acquisition unit is used to preprocess the mobile phone signaling data in the multi-source data, and to match the preprocessed mobile phone signaling data of each base station with the multiple street units to obtain the user's spatial activity trajectory. A computing unit is used to calculate the attribute representation of each of the plurality of block units, wherein the attribute representation includes at least one of spectral features, texture features, geometric morphology features and economic semantic features; The construction unit is used to construct the graph structure data based on the initial functional area distribution, the spatial activity trajectory, and the attribute representation.

8. The apparatus according to claim 7, characterized in that, The computing unit is specifically used for: Acquire remote sensing images, and extract the mean and standard deviation of each band from the remote sensing images, and obtain the spectral features based on the mean and standard deviation of each band; Obtain the gray-level co-occurrence matrix of the remote sensing image, and obtain the texture features based on the gray-level co-occurrence matrix; Obtain building outline data, and extract the geometric features from the building outline data; A POI corpus is constructed using a preset algorithm. The POI corpus is then input into a preset model to obtain the semantic embeddings of POI categories. The semantic embeddings of the POI categories are then weighted and aggregated to obtain the economic semantic features.

9. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, the processor executing the program to implement the urban functional area identification method as described in any one of claims 1-5.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the urban functional area identification method as described in any one of claims 1-5.

Images