Site selection method for urban emergency shelters based on VGAE-RF model

Abstract

The application provides a city emergency shelter site selection method and device based on a VGAE-RF model, equipment and a medium, wherein the method comprises the following steps: acquiring multi-source spatial data in an open data source; performing spatial subdivision on a target region based on a regular polygon division method to obtain spatial grid cells; extracting the explicit features of city structures, geographical environments and social attributes in each spatial grid cell one by one based on the multi-source spatial data, and constructing an explicit feature matrix; constructing a global space topology graph, sampling the implicit features coupled between the spatial grid cells in combination with a VGAE model, and constructing an implicit feature matrix; and fusing the explicit features and the implicit features, and performing site selection prediction of city emergency shelters based on an RF model, so as to solve the problem of low accuracy of site selection results in predicting the positions of city emergency shelters.

Description

Technical Field

[0001] The present invention relates to the field of geographic information technology, and in particular to a method, device, equipment and medium for selecting a site for an urban emergency shelter based on a VGAE-RF model. Background Art

[0002] With the acceleration of urbanization, the urban spatial structure shows a trend of multi-centerization, dense and high-rise buildings, and concentrated and diversified population. The corresponding emergency shelters are key facilities for responding to urban emergencies, and their importance is becoming increasingly prominent. However, compared with the rapid expansion and complexity of cities, the planning and construction of emergency shelters often lag behind.

[0003] Traditional site selection methods within a multi-criteria decision-making framework primarily rely on pre-established candidate locations and select sites based on preset evaluation criteria. This approach relies heavily on the subjective experience of decision-makers, simplifying the high-level and unstructured nature of regions. However, with the development of smart cities, the volume of multi-source, heterogeneous urban spatial data has rapidly increased, providing rich data support for theoretical research on urban site selection. Compared with traditional site selection methods, artificial intelligence-based site selection methods driven by multi-source data are more capable of overcoming existing limitations. Some studies have adopted an end-to-end analytical approach, using data such as building footprints, user visit trajectories, and urban points of interest to design complex site selection models for applications in commerce, healthcare, transportation, and other fields. However, these multi-source data-driven site selection studies often treat candidate points or geographic units as isolated objects, ignoring the influence of neighboring information within the complex geographic space.

[0004] Existing GNNs (Graph Neural Networks), with their inherent topological sensitivity and powerful node feature fusion capabilities, can understand and utilize proximity information in geographic space at a higher dimension and deeper level, effectively revealing and exploiting these underlying spatial connections. Some studies have employed GNNs to process geographic relationships, but most rely on pre-defined points in specific scenarios to construct spatial graph convolutional models, or predict the attractiveness of facilities such as stores and hospitals based on user visit trajectories or public transportation services. Such constructed spatial attribute graphs produce inconsistent spatial meanings, preventing the full consideration of potentially exploitable spaces. These methods can adapt to large-scale and complex computations, significantly improving data processing efficiency and expanding the scope of site selection while avoiding inaccurate site selection results caused by overly subjective approaches. However, in a geographical environment characterized by spatial correlation and heterogeneity, these methods isolate site selection objectives as separate entities, which is clearly undesirable.

[0005] Therefore, it is urgent to propose a method for site selection of urban emergency shelters based on the VGAE-RF model to solve the technical problem of low accuracy of site selection results for predicting the location of urban emergency shelters in the existing technology. Summary of the Invention

[0006] In order to overcome the problems existing in the related art, the present disclosure provides a method and device for urban emergency shelter site selection based on the VGAE-RF model to solve the technical problem of low accuracy in the site selection results of predicting the location of urban emergency shelters in the related art.

[0007] One or more embodiments of this specification provide a method for selecting an urban emergency shelter site based on a VGAE-RF model.

[0008] Access multi-source spatial data from open data sources;

[0009] The target area is spatially divided based on the regular polygon partitioning method to obtain spatial grid units;

[0010] Extracting the explicit features of the urban structure, geographical environment, and social attributes within each spatial grid unit one by one based on the multi-source spatial data, and constructing an explicit feature matrix;

[0011] Constructing a global spatial topology map, combining it with the VGAE model, sampling the implicit features of the coupling between the spatial grid units, and constructing an implicit feature matrix;

[0012] The explicit feature matrix and the implicit feature matrix are integrated, and the location prediction of urban emergency shelters is performed based on the RF model.

[0013] Preferably, the dominant features of the urban structure include road network density features and building features. The step of extracting the dominant features of the urban structure within each spatial grid unit one by one based on the multi-source spatial data and constructing a dominant feature matrix comprises the following steps:

[0014] The road network density is comprehensively calculated by the ratio of the total length of the four types of roads in the spatial grid unit to the area of ​​the spatial grid unit and the road network density is linearly normalized to construct a road network density feature matrix;

[0015] The building density is determined by the number of buildings per unit area, the average building height is determined by dividing the total building height by the number of buildings, and the building density and the average building height are linearly normalized to construct a building feature matrix.

[0016] Preferably, the explicit features of the geographical environment include average elevation, slope mean, slope standard deviation, and elevation standard deviation, and the step of constructing an explicit feature matrix based on the explicit features of the geographical environment in each spatial grid cell one by one includes the following steps:

[0017] The average elevation, the slope mean, the slope standard deviation, and the elevation standard deviation are linearly normalized by column to construct a geographical environment feature matrix.

[0018] Preferably, the explicit features of the social attribute include population density and POI density, and the step of constructing an explicit feature matrix based on the explicit features of the social attribute in each spatial grid cell one by one includes the following steps:

[0019] The population density is calculated using the ratio of the total number of populations in the spatial grid cell to the area of the spatial grid cell to construct a population density feature matrix.

[0020] A POI number matrix is constructed, a POI density matrix is calculated, a reclassification is extracted by normalizing the POI density matrix, and a POI density feature matrix is constructed.

[0021] Preferably, the step of constructing a global spatial topology graph, combining a VGAE model, sampling the implicit features between the spatial grid cells, and constructing an implicit feature matrix includes the following steps:

[0022] A global spatial topology graph is constructed according to the explicit features in the spatial grid cell and the adjacency relationship of the regular polygon.

[0023] The explicit features of the urban structure, the geographical environment, and the social attribute are constructed as a feature matrix of the node.

[0024] In the VGAE model, the explicit features of each spatial grid cell are learned by an encoder including two layers of graph convolutional neural networks to obtain the mean and variance of the low-dimensional spatial representation.

[0025] The implicit features are sampled in the low-dimensional space by a decoder, and the spatial topology graph is reconstructed by an inner product operation to construct an implicit feature matrix.

[0026] Preferably, the loss function of the VGAE-RF model includes two parts: a weighted cross-entropy loss function and a KL divergence of the implicit spatial distribution and the standard normal distribution.

[0027] One or more embodiments of the present specification provide a VGAE-RF model-based urban emergency shelter site selection device, comprising:

[0028] A data generation module is configured to obtain multi-source spatial data in an open data source.

[0029] The spatial partitioning module is used to spatially partition the target area based on the regular polygon partitioning method to obtain spatial grid units;

[0030] An explicit feature extraction module is used to extract the explicit features of the urban structure, geographical environment and social attributes within each spatial grid unit one by one based on the multi-source spatial data, and construct an explicit feature matrix;

[0031] The implicit feature extraction module is used to construct a global spatial topology map, sample the implicit features of the coupling between the spatial grid units in combination with the VGAE model, and construct an implicit feature matrix;

[0032] The prediction module is used to fuse the explicit feature matrix and the implicit feature matrix and perform site selection prediction of urban emergency shelters based on the RF model.

[0033] Preferably, the latent feature extraction module includes:

[0034] A topology map construction unit, configured to construct a global spatial topology map based on the dominant features within the spatial grid cells and the adjacency relationship of the regular polygons;

[0035] A matrix construction unit is used to construct the explicit characteristics of urban structure, geographical environment and social attributes into a feature matrix of nodes;

[0036] The low-dimensional spatial operation unit is used in the VGAE model to learn the explicit features of each spatial grid cell through an encoder containing a two-layer graph convolutional neural network to obtain the mean and variance of the low-dimensional spatial representation;

[0037] The latent feature sampling unit samples latent features in the low-dimensional space through a decoder, reconstructs a spatial topology map through an inner product operation method, and constructs a latent feature matrix.

[0038] One or more embodiments of this specification provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the urban emergency shelter site selection method based on the VGAE-RF model as described above is implemented.

[0039] One or more embodiments of this specification provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-mentioned method for selecting an urban emergency shelter site based on the VGAE-RF model.

[0040] The method, device, equipment and medium for site selection of urban emergency shelter based on a VGAE-RF model are provided in the disclosure, and the advantages are as follows: multi-source spatial data in an open data source is acquired to obtain a data set that can support theoretical research on urban site selection; a target region is spatially divided based on a regular polygon division method to obtain spatial grid units, and the experimental area range is fully covered and seamlessly spatially divided; explicit features inside each spatial grid unit are extracted one by one based on the multi-source spatial data; a global spatial topology graph is constructed, and VGAE model is combined to sample implicit features coupled between the spatial grid units, construct an implicit feature matrix, and realize sampling of the implicit feature information coupled between the spatial grid units and construction of the feature matrix by integrating more comprehensive influencing factor analysis; the explicit feature matrix and the implicit feature matrix are fused, and site selection prediction of the urban emergency shelter is performed based on the RF model, so that the site selection result is more uniform, the coverage range is wider, the reliability and accuracy of the site selection result are improved, and over-concentration or sparse distribution of the emergency shelter can be avoided to a greater extent, thereby improving the overall spatial suitability. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the one or more embodiments of the present disclosure or the prior art, the drawings needed to be used in the embodiment or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments described in the present disclosure, and other drawings can be obtained by those skilled in the art without any creative labor.

[0042] Figure 1 A flowchart of a method for site selection of urban emergency shelter based on a VGAE-RF model is provided for one or more embodiments of the present disclosure.

[0043] Figure 2 A general framework diagram of a method for site selection of urban emergency shelter based on a VGAE-RF model is provided for one or more embodiments of the present disclosure.

[0044] Figure 3 A schematic diagram of a graph adjacency matrix A and a feature matrix X of a node is provided for one or more embodiments of the present disclosure.

[0045] Figure 4 A schematic diagram of aggregated neighborhood information of an autoencoder is provided for one or more embodiments of the present disclosure.

[0046] Figure 5 A curve graph of changes in performance of a VGAE-RF under different fusion parameter α values is provided for one or more embodiments of the present disclosure.

[0047] Figure 6The ROC characteristic curve diagram provided for one or more embodiments of the present specification;

[0048] Figure 7 The VGAE-RF refuge site selection result comparison diagram with existing shelters provided for one or more embodiments of the present specification;

[0049] Figure 8 The structural diagram of a VGAE-RF model-based urban emergency shelter site selection device provided for one or more embodiments of the present specification;

[0050] Figure 9 The structural diagram of a computer device provided for one or more embodiments of the present specification. DETAILED DESCRIPTION

[0051] In order to make the person skilled in the art better understand the technical solutions in one or more embodiments of the present specification, the technical solutions in one or more embodiments of the present specification will be described clearly and completely in conjunction with the drawings in one or more embodiments of the present specification. Obviously, the described embodiments are only part of the embodiments of the present specification, not all embodiments. Based on one or more embodiments of the present specification, all other embodiments obtained by those skilled in the art without creative labor should fall within the protection scope of the present invention.

[0052] The present invention will be described in detail below in conjunction with the specific implementation manner and the drawings in the specification.

[0053] Method embodiment

[0054] According to the embodiments of the present application, a VGAE (Variational Graph Auto Encoders, Variational Graph Auto Encoders)-RF (Random Forest, Random Forest) model-based urban emergency shelter site selection method is provided, as shown in Figure 1 The flowchart of the VGAE-RF model-based urban emergency shelter site selection method provided for the present embodiment, according to the VGAE-RF model-based urban emergency shelter site selection method of the embodiments of the present application, includes:

[0055] Step S10, acquiring multi-source spatial data in open data sources.

[0056] In step S10, by using the combination of meta-search and deep search engines, according to different types and different structures of data characteristics, different data collection strategies are adopted to collect DEM (Digital Elevation Model) data, POI (Point of Interest), emergency shelter location, population network, road network and administrative division, and other information types related to emergency shelter site selection in the multi-source network.

[0057] In a specific embodiment, the specific data collection strategy mainly includes:

[0058] (1) For structured POI data, considering that it usually exists in the form of map service content, it is often necessary to use python scripts to download in batches from different open map platforms.

[0059] (2) For data published in the form of semi-structured data packets, engine search is adopted combined with manual discrimination collection, such as geographic spatial data cloud and national Qinghai-Tibet Plateau scientific data center.

[0060] (3) For unstructured statistical information, such as seven population and shelter attributes, crawler technology is used to obtain and further process semi-structured information to supplement or verify structured geographic data.

[0061] Due to the differences in spatial reference and coordinate coding of heterogeneous geographic data, accompanied by data missing and redundancy problems, the geometry, attributes and spatial relationships of geographic entities are inconsistent. Before feature construction, the multi-source heterogeneous data needs to be preprocessed, including the following steps:

[0062] Step 101, data cleaning, coordinate transformation, geometry correction, reorganization and missing data checking are performed on the original heterogeneous data collected from the network to ensure the consistency of geographic data expression and provide a basis for data fusion;

[0063] Step 102, consistent processing of spatial relationships of multi-source data, such as complementary correction of population statistics data and raster data, road network topology repair, and alignment of emergency shelter locations;

[0064] Step 103, verification and evaluation of data processing results and data quality reliability, completeness and precision, and segmentation of data units using the boundary of the study area to provide a basis for feature extraction and correlation analysis.

[0065] Step S20, based on the division method of regular polygon, the target area is spatially divided to obtain spatial grid units.

[0066] Specifically, there are three main methods for dividing plane space into regular polygons: triangles, squares, and hexagons. The adjacency relationships generated by triangle and square divisions may lead to the connectivity paradox. Regular hexagons, on the other hand, offer several significant advantages in geospatial analysis:

[0067] (1) Uniform adjacency, effectively eliminating the ambiguity in determining the adjacency relationship of spatial grid cells;

[0068] (2) A higher perimeter-to-area ratio can more effectively reduce edge effects and the resulting orientation bias and sampling bias;

[0069] (3) The distance between edge points and center points is balanced, which helps to more accurately fit the multi-directional and multi-dimensional characteristics of the spatial service range. This embodiment takes advantage of these advantages, adopts regular hexagons as the basic unit for subdividing spatial grids, and uses the neighborhood relationships determined by them to construct a spatial topology map.

[0070] In selecting the scale of the spatial grid cells, this example was based on research into the time and walking distance required for residents to reach shelters in disaster scenarios, as well as reference to relevant domestic and international literature and construction standards. The final scale was 500 meters. At this scale, the interior of the regular hexagonal grid is considered uniform, and the characteristic factors of each subdivision are associated with its corresponding unique identifier, facilitating detailed and accurate spatial analysis.

[0071] Step S30: Based on the multi-source spatial data, the explicit features of the urban structure, geographic environment, and social attributes within each spatial grid cell are extracted one by one to construct an explicit feature matrix. The factors influencing the selection of urban emergency shelters are categorized into three core dimensions: urban structure, geographic environment, and social attributes. These dimensions involve different data types and representations, and different methods are used to extract explicit features, thereby constructing a comprehensive explicit feature matrix.

[0072] Step S40: construct a global spatial topology map, combine it with the VGAE model, sample the implicit features of the coupling between spatial grid units, and construct an implicit feature matrix.

[0073] Step S50: Fusing the explicit feature matrix and the implicit feature matrix, and performing site selection prediction for urban emergency shelters based on the RF model.

[0074] In step S50, since explicit features contain more intuitive attribute information such as urban structure, geographical environment and social attributes, they have a dominant influence on the location of emergency shelters. However, explicit features are limited to the defined scale, while implicit features contain multi-order neighbor aggregation semantic information determined by the number of convolution layers. In the research and application of many deep learning algorithms, fusing features at different levels is an important means to improve the performance of downstream tasks. Therefore, on the basis of constructing the explicit features X of the geographic unit map nodes, combined with the implicit features Z of the aggregated second-order neighborhood images mined by VGAE, the two feature attributes are weightedly fused. The specific calculation is shown as follows:

[0075]

[0076] in, It represents the matrix after the fusion of explicit and implicit features, α is the weight coefficient, X is the explicit feature matrix, and Z is the implicit feature matrix.

[0077] Taking into account the characteristics of the dimension, nonlinear data and other factors, and by comparing and analyzing a variety of classic machine learning models and methods, this paper selects the random forest model for predictive analysis of emergency shelter site selection. The RF model is a multi-decision tree ensemble model that has good accuracy and generalization in solving high-dimensional nonlinear classification problems. The model constructs a set of base evaluators {h1(X),h2(X),...,h k (X)}, each round of training uses the Bootstrap method to extract samples from the original data set with replacement, and the final prediction result is obtained by comprehensive voting of multiple rounds of decision trees:

[0078]

[0079] Among them, H(x) is the final classification result output by the model, and I(·) is the indicative function used to judge the classification results of each decision tree h. i Is the output of (X) equal to the given target variable y?

[0080] Existing emergency shelters are the result of careful consideration by decision-makers. The labels used in this study are based on shelter data from POIs and publicly available government emergency shelter information. During the prediction phase, we used ADASYN to address the imbalance of fused feature data and split the sample size into a 7:3 ratio for training and validation.

[0081] like Figure 2 , which is a general framework diagram of a method for selecting an urban emergency shelter site based on a VGAE-RF model provided in this embodiment.

[0082] The method provided by the embodiment obtains a data set capable of supporting theoretical research on city site selection by acquiring multi-source spatial data in an open data source; a target region is spatially divided based on a regular polygon division method to obtain spatial grid cells, and the experimental area range is fully covered and seamlessly spatially divided; explicit features inside each spatial grid cell are extracted one by one based on the multi-source spatial data; a global spatial topology graph is constructed, and the implicit features coupled between the spatial grid cells are sampled in combination with a VGAE model to construct an implicit feature matrix, and through the analysis of more comprehensive influencing factors, the sampling of the implicit feature information coupled between the spatial grid cells and the construction of the feature matrix are realized; the explicit feature matrix and the implicit feature matrix are fused, and site selection prediction of the city emergency shelter is performed based on an RF model, so that the site selection result is more uniform and has a wider coverage, and the reliability and accuracy of the site selection result are improved, and the over-concentration or sparse distribution of the emergency shelters can be avoided to a greater extent, and thus the overall spatial suitability is improved.

[0083] In one embodiment, the explicit features of the city structure in step S30 include road network density features and building features, and the extracting of the explicit features of the city structure inside each spatial grid cell based on the multi-source spatial data and the constructing of the explicit feature matrix include the following steps.

[0084] S301, the road network density is comprehensively calculated by the ratio of the total length of the four types of roads in the spatial grid cell to the area of the spatial grid cell, and linear normalization processing is performed on the road network density to construct a road network density feature matrix; the building density is determined by the number of buildings per unit area, the average building height is determined by dividing the total building height by the number of buildings, and linear normalization processing is performed on the building density and the average building height to construct a building feature matrix.

[0085] (1) Road network density

[0086] The four types of roads, i.e., city main roads, city secondary roads, city branch roads, elevated roads and expressways, are selected as data for calculating the road network density, and the road network density is comprehensively calculated by the ratio of the total length of the four types of roads in the spatial grid cell to the area of the spatial grid cell, and linear normalization processing is performed on the road network density.

[0087] Specifically, the road network density formula is as follows:

[0088]

[0089] D = [D1, D2,..., D i ] T ;

[0090] wherein D i represents the road network density of the i-th spatial grid cell without normalization, l ijrepresents the length of the j-type road in the unit, S represents the area of ​​the spatial grid unit, and D represents the D of all spatial grid units. i The column vector is composed of the values ​​and transposed, and T represents the transpose operation of the vector or matrix.

[0091] Next, the road network density is linearly normalized:

[0092]

[0093] Among them, D min represents the minimum value in D, D max represents the maximum value in D, X Road It represents the linearly normalized value of the road network density in all spatial grid cells.

[0094] (2) Architectural features

[0095] To quantify building concentration and floor height characteristics, we first construct a building attribute matrix B = [ID, H, BN], where ID is the identifier of spatial grid cell i, H represents the total building height within the spatial grid cell, and BN represents the number of buildings within the spatial grid cell. Building density is determined by the number of buildings per unit area, and average building height is determined by dividing the total building height by the number of buildings. Both building density and average building height are linearly normalized.

[0096] Specifically, the building density calculation formula is as follows:

[0097]

[0098] Among them, BD i represents the building density of geographic unit i, B ki3 represents the number of buildings in spatial grid cell i, and S is the area of ​​spatial grid cell i.

[0099] The formula for calculating the average building height is as follows:

[0100]

[0101] Among them, AH i represents the average building height of geographic unit i, B ki2 represents the total building height in geographic unit i.

[0102] Next, in order to eliminate the influence of dimension, the features are linearly normalized so that their values ​​are mapped between 0 and 1:

[0103] BD=[BD1,BD2,...,BD i ] T ;

[0104] AH=[AH1,AH2,...,AH i ] T ;

[0105]

[0106] Among them, BD and AH are column vectors, including the building density BD of spatial grid cell i i and average building height AH i , X Building is the linearly normalized building feature matrix, BD min is the minimum value of vector BD, BD max is the maximum value of vector BD, AH min is the minimum value of vector AH, AH max is the maximum value of vector AH.

[0107] The method provided in this embodiment can obtain complete and high-precision urban structure data features by extracting the dominant features of road network density and building features in the urban structure and performing data normalization processing, thereby providing data support for association analysis.

[0108] In one embodiment, in step S30, the geographical environment includes the average altitude, the mean slope, the standard deviation of the slope, and the standard deviation of the elevation. Based on the multi-source spatial data, the dominant features of the geographical environment within each spatial grid cell are extracted one by one to construct a dominant feature matrix, which includes the following steps:

[0109] S302: Perform linear normalization on the average altitude, slope mean, slope standard deviation, and elevation standard deviation by column to construct a geographic environment feature matrix.

[0110] In the planning of emergency shelters, geographical environmental factors such as slope and altitude are crucial. Slope affects the accessibility of the shelter, while altitude is closely related to the shelter's resistance to natural disasters such as floods, safety and stability in emergency situations, etc. This embodiment constructs geographical environmental characteristics based on pre-processed DEM data from four aspects: average altitude, mean slope, slope standard deviation, and elevation standard deviation. For any spatial grid cell, the slope information is expressed as:

[0111] Slop i =[s1,s2,...,s N ];

[0112] Among them, s N is the slope of each sampling point, and N is the number of sampling points in the spatial grid unit.

[0113] Elevation information is represented as:

[0114] Hi =[h1,h2,...,h N ];

[0115] Among them, h N The elevation of each sampling point.

[0116] The mean altitude is expressed as:

[0117]

[0118] The mean slope is expressed as:

[0119]

[0120] Specifically, the slope standard deviation can indicate the richness of regional terrain types, while the elevation standard deviation can measure the susceptibility to natural disasters such as floods. The slope standard deviation is calculated as follows:

[0121]

[0122] The elevation standard deviation is calculated as follows:

[0123]

[0124] The method provided in this embodiment can obtain complete and high-precision geographic environment data features by extracting the dominant features of the average altitude, slope mean, slope standard deviation, and elevation standard deviation in the geographic environment and performing data normalization processing, thereby providing data support for association analysis.

[0125] In one embodiment, the explicit features of the social attributes in step S30 include population density and POI density. Extracting the explicit features of the social attributes within each spatial grid cell one by one based on the multi-source spatial data and constructing an explicit feature matrix includes the following steps:

[0126] S303. Calculate the population density using the ratio of the total population of the spatial grid unit to the area of ​​the spatial grid unit, and construct a population density feature matrix; construct a POI quantity matrix, calculate the POI density matrix, and construct a POI density feature matrix by normalizing, extracting, and reclassifying the POI density matrix.

[0127] (1) Population density

[0128] Based on the grid population data corrected in the seventh national census, the population density is calculated using the ratio of the total population of the spatial grid unit to the area of ​​the spatial grid unit. The population density formula is as follows:

[0129]

[0130] Among them, Popi represents the population density in the i-th spatial grid cell, PN i is the population in spatial grid cell i, and S is the area of ​​geographic cell i.

[0131] (2) POI density

[0132] POI refers to the location points that the public pays attention to within a specific spatial range. These points can reflect the population activity footprint in the area to a certain extent and are one of the factors affecting the site selection of emergency shelters. However, due to the diversity of POI types and uneven data quality, it is necessary to process and select POI types related to emergency shelters when constructing features. This embodiment reclassifies POIs within a geographical unit according to the "Land Use Classification Standard (GB / T 21010-2017)" and relevant emergency field literature. The eight types include residential services, catering and commercial services, industrial services, government functions, companies and enterprises, dangerous evacuation, key evacuation, and park green space. The classification details are shown in Table 1.

[0133] Table 1. POI reclassification information

[0134]

[0135] After reclassifying the POIs, the ratio of the number of facilities in each spatial grid unit to the grid area of ​​each POI is calculated to obtain the density of each type of POI. The calculation is as follows:

[0136] M=[N L1 ,N L2 ,...,N Lj ];

[0137]

[0138] Among them, M represents the number matrix of various POI facilities in each spatial grid unit, L j represents the j-th type of POI, N Lj Represents L in spatial grid cell i j Number of POI-like facilities, PD ij Represents L in spatial grid cell i j Class POI density, M ij Represents L in spatial grid cell i j The number of POI-like facilities, S represents the area of ​​each spatial grid unit.

[0139] Through linear normalization, the density of various POIs PD ij The resulting density matrix PD is mapped between 0 and 1:

[0140] The calculation is as follows:

[0141]

[0142] Where: X poi (i, j) is the L in the normalized spatial grid cell i j POI-like facility density, PDmin j is the minimum value of each column in the matrix PD, PDmax j is the minimum value of each column in the matrix PD.

[0143] The method provided in this embodiment can obtain complete and high-precision social attribute data features by extracting the explicit features of population density and POI density in social attributes and performing data normalization processing, thereby providing data support for association analysis.

[0144] In one embodiment, to improve the site selection of emergency shelters, it is necessary to consider not only the explicit features within the spatial grid cells, but also the potential impact of the spatial grid cells that are interconnected with them. Therefore, in step S40, a global spatial topology map is constructed, and the implicit features of the coupling between spatial grid cells are sampled in combination with the VGAE model to construct an implicit feature matrix, including the following steps:

[0145] S401. Utilize the internal explicit features of spatial grid cells and their corresponding regular polygon-regular hexagon adjacency relationships to construct a global spatial topology graph G = (V, E) to represent the surface connections of discrete geographic units in space.

[0146] S402. Construct the explicit characteristics of urban structure, geographical environment and social attributes into a feature matrix of nodes.

[0147] In the spatial topology graph G, V represents the node set corresponding to each spatial unit, and E is the set of edges, where the element e ij =(v i ,v j ) represents node v i With v j There are adjacencies between them. Figure 3 As shown in the figure, it is a schematic diagram of the graph adjacency matrix A and the node feature matrix X provided in this embodiment, v i ∈V represents the nodes in the graph, A represents the adjacency matrix of the nodes, if e ij ∈E then a ij =1 means that spatial units i and j are adjacent, otherwise a ij = 0. X represents the feature matrix of the node, each node v i The property x i ∈X is characterized by the explicit features of the corresponding spatial grid cells. That is, the features of the three core dimensions of geographical environment, urban structure, and social attributes are constructed as the feature matrix X of the node, forming the semantic information of the graph:

[0148] X=X E ∈R N×F ;

[0149] Where N is the number of nodes in the spatial topology graph, and F is the explicit feature dimension.

[0150] S403. In the VGAE model, the explicit features of each spatial grid unit are learned through an encoder containing a two-layer graph convolutional neural network to obtain the mean and variance of the low-dimensional spatial representation.

[0151] Spatial topology graphs connect discrete geographic units to each other, and graph neural networks can sample the potential semantic features between interconnected nodes when processing graph data. GCN (Graph Convolutional Networks) is popular among researchers because of its ability to aggregate neighborhood information. However, the semi-supervised GCN model will cause the overall semantics of the graph to be destroyed during the learning and training phase, and it also lacks the ability to capture the latent spatial patterns in the graph data. Therefore, we introduce VGAE, which combines the advantages of GAE (Graph Auto-encoders) and VAE (Variational Auto-encoders) to capture complex patterns in node representations more deeply. Specifically, the VGAE model first uses an encoder consisting of two layers of graph convolutional neural networks to encode the explicit features x of each geographic unit. i Learn to obtain the mean μ and variance σ of the low-dimensional vector representation. The encoder consists of a two-layer graph convolutional neural network, and the graph convolution operation is defined as:

[0152]

[0153] in, is the Laplace matrix of the adjacency matrix A, W0 and W1 are the weight parameters to be updated.

[0154] In this way, the two-layer GCN effectively aggregates the second-order neighborhood information to learn the latent spatial representation of the target geographic unit. Figure 4 FIG. 1 is a schematic diagram of aggregating neighborhood information by the autoencoder provided in this embodiment.

[0155] S404: Sample the latent feature Z in the low-dimensional space through the decoder, reconstruct the spatial topology A' through the inner product operation method, and construct the latent feature matrix.

[0156] The decoder is modeled as follows:

[0157]

[0158] in,

[0159] The latent feature representations obtained by VGAE are different from those obtained by traditional GCN. In traditional GCN, latent feature representations are usually learned directly from the input data. In VGAE, these representations are sampled from a multi-dimensional Gaussian distribution. The advantage of this is that it can better capture and represent the inherent structure and complexity of the data. In this process, the mean vector and covariance matrix of the multi-dimensional Gaussian distribution are learned by the second layer of the encoder GCN. The specific expression is:

[0160]

[0161]

[0162] in, and are the weight matrices of the second layer of GCN, and σ(·) is the sigmoid function.

[0163] In one embodiment, the loss function of the VGAE-RF model includes two parts: a weighted cross entropy loss function and a KL divergence between the latent space distribution and the standard normal distribution.

[0164] In a specific embodiment, the loss function of the VGAE-RF model consists of two parts: one is the weighted cross entropy loss function, which is used to measure the difference between the original image and the reconstructed image. It makes the reconstructed image as close to the original image as possible to ensure that the model can accurately capture and learn the structure of the data; the other part is the KL divergence between the latent space distribution and the standard normal distribution, which ensures the stability and generalization ability of the model.

[0165] The loss function can be calculated as:

[0166] L=E q(Z|X,A) [logp(A|Z)]-KL[q(Z|X,A)‖p(Z)];

[0167] In a specific example, Beijing was used as an experimental area. Open-source spatial data was collected from multiple sources, and explicit and implicit features were extracted. The prediction method described in this invention was then applied to verify the effectiveness of the proposed emergency shelter site selection method. The predicted value of each node reflects the overall characteristics of the city's geographic unit, thereby indicating whether the area is an optimal location for a shelter.

[0168] The growing volume of multi-source geospatial data makes it possible to explore the "human-land" coupling relationship from a micro and dynamic perspective, and promotes urban planning and design, providing important decision-making information for urban planning. In the experiment, multi-source geospatial data were collected through open data sources, and the data set was constructed according to the above-mentioned characteristics. The specific data acquisition time and source have been listed in detail in the experiment to ensure the traceability and accuracy of the data.

[0169] Table 2, data description and source

[0170]

[0171]

[0172] For model reliability, the precision (Precision), recall (Recall) and F1 score based on the confusion matrix are selected as one of the evaluation reliability indicators in the present application:

[0173]

[0174]

[0175]

[0176] Wherein, TP is the number of true classes, FP is the number of false positive classes, and FN is the number of false negative classes. In addition, in the face of the entire city perspective, the number of emergency shelters is much smaller than that of other public service facilities, resulting in a class imbalance phenomenon in the construction of the final fusion feature matrix However, the ROC curve (Receiver Operating Characteristic) and AUC value (Area Under roc Curve) can not only measure the performance of the classifier well, but also avoid the evaluation deviation problem caused by the distribution imbalance of the sample. The ROC curve depicts the performance of the classifier at different thresholds by taking the true positive rate (True Positive Rate, TPR) and false positive rate (False Positive Rate, FPR) as the horizontal and vertical coordinates, and the AUC value is the area covered by the ROC curve, between 0 and 1, the larger the value, the better the model. TPR and FPR can be calculated according to the following formula:

[0177]

[0178]

[0179] Wherein, TN is the number of true negative classes.

[0180] The unsupervised learning characteristics of the VGAE-RF model in this embodiment enable us to mine the potential features in the city spatial topology without considering the division of the data set, avoiding the incomplete learning of potential features caused by data splitting. For the mining of node hidden features, a variational graph autoencoder (VGAE) containing two layers of graph convolutional networks (GCN) is constructed as an encoder in the present application. The input of the model is the node features, the output dimension of the hidden layer is set to 16, the initial learning rate is set to 0.01, the weight decay is 5e-4, and the dropout rate is 0.00. The table shows the hyperparameters in the VGAE.

[0181] Table 3, VGAE model hyperparameters

[0182]

[0183] Before the feature fusion stage, the hidden features learned by the VGAE-RF need to be normalized to avoid the influence of negative numbers on the original features. In order to select the best feature fusion weight parameter a, a is sampled from 0 to 1 with a step size of 0.01 for prediction. In the experiment, the precision, recall, and F1 score of the prediction results of different weight feature fusion are obtained, and the "evaluation-weight" curve diagram is as shown in Figure 5 Fig. 1, when a = 0.85, the F1 score (0.9018) in the evaluation result is the best.

[0184] In order to further verify the generalization ability and reliability of the model, 10-fold cross-validation is used to measure the performance of the model in the experiment, and the average values of precision (Precision), recall (Recall), and F1 score are calculated. In this process, a variety of baseline models are compared with the VGAE-RF model, including random forest, support vector machine (SVM), K- nearest neighbor (KNN), and multi-layer perceptron (MLP). Then, these baseline models are used instead of the random forest-based classifier in the original method to analyze the performance of the fused feature matrix in different base classifiers. Finally, the VGAE-RF-NF model without fused original features is also compared with the VGAE-RF model. Table 4 shows the evaluation scores of different models in the above experiment.

[0185] Table 4, comparison of model evaluation scores

[0186]

[0187] The results show that the VGAE-RF model achieves the highest precision (0.9112) and F1 score (0.9018), and its recall rate (0.8904) is relatively close to that of VGAE-MLP. In summary, (1) among all the methods, the non-neural network methods (RF, SVM, KNN) that ignore the influence of spatial grid unit coupling and directly predict are the least reliable. On the one hand, it is difficult for simple models to learn the complex feature relationships of geographic units. On the other hand, after the area is subdivided into surfaces, the model receptive field is only in a single-scale geographic unit and cannot capture the potential semantic relationship between units. (2) Among the spatial prediction methods that use spatial topology maps to fuse explicit and implicit features, the corresponding simple baseline models have a certain degree of improvement in performance. This shows that the potential influence of spatial grid units is a factor that cannot be ignored, and the VGAE-RF model can well capture this potential spatial relationship. (3) Compared with the results before and after the fusion of explicit and implicit features, the method without fusion of original features (VGAE-RF-NF) has better performance than the non-neural network method, but the fused VGAE-RF model is more reliable. That is, aggregating the second-order neighbor node features through the VGAE-RF model may cause some information loss of the original features. This problem can be well solved by feature fusion technology.

[0188] The effectiveness of the method of this embodiment is illustrated below through a specific case.

[0189] In this case, Figure 6 As shown in the ROC characteristic curve diagram provided by this embodiment, the ROC characteristic curve of the VGAE-RF model obtained by calculating the false positive rate (FPR) and the true positive rate (TPR) is significantly tilted to the upper left. This characteristic is consistent with the performance of an ideal classifier, indicating that the model can maintain good performance under conditions of low false positive rate and high true positive rate, which not only shows that the model has reliability in predicting imbalanced data, but also reveals its high classification accuracy and robustness. In addition, the area under the curve (AUC) is 0.979, which further confirms the reliability and effectiveness of the VGAE-RF model, because in the evaluation of the classification model, the closer the AUC value is to 1, the better the model performs in distinguishing positive and negative samples.

[0190] This example sorts the predicted probabilities of VGAE-RF and uses the spatial distribution of population and the area of ​​the 16 districts under Beijing as weights to select 600 candidate areas from all prediction units. Figure 7As shown, the VGAE-RF model shelter site selection results provided in this embodiment are compared with existing shelters. The VGAE-RF model site selection results are compared with existing emergency shelters. The model site selection distribution is more uniform and covers a wider range than existing emergency shelters. It can avoid excessive concentration or sparse distribution of shelters to a greater extent, thereby improving the overall spatial suitability. At the same time, the high-density areas of the model site selection have a high degree of overlap with densely populated areas, which helps to achieve faster evacuation and rescue in emergencies and reduce the potential risk of casualties. Moreover, the newly added points are not only closer to important infrastructure such as medical care and transportation, but also far away from potential disaster-sensitive areas, such as gas stations, liquefied gas stations and other high-risk facilities. This is conducive to providing timely rescue and material support when a disaster occurs, ensuring the safety and availability of emergency shelters when a disaster occurs.

[0191] Device embodiment

[0192] According to an embodiment of the present invention, a device for selecting an urban emergency shelter site based on a VGAE-RF model is provided. Figure 8 FIG. 1 is a schematic diagram of the structure of the urban emergency shelter site selection device based on the VGAE-RF model provided in this embodiment. The urban emergency shelter site selection device based on the VGAE-RF model according to the embodiment of the present invention includes:

[0193] The data generation module 610 is used to obtain multi-source spatial data from open data sources and pre-process the multi-source spatial data.

[0194] The space division module 620 is used to perform spatial segmentation on the target area based on a regular polygon division method to obtain spatial grid units.

[0195] The explicit feature extraction module 630 is used to extract the explicit features of the urban structure, geographical environment and social attributes within each spatial grid unit one by one based on multi-source spatial data, and construct an explicit feature matrix.

[0196] The implicit feature extraction module 640 is used to construct a global spatial topology map, combine the VGAE model, sample the implicit features of the coupling between spatial grid units, and construct an implicit feature matrix.

[0197] The prediction module 650 is used to fuse the explicit feature matrix and the implicit feature matrix and perform site selection prediction of urban emergency shelters based on the RF model.

[0198] The device provided by the embodiment obtains multi-source spatial data in an open data source through the data generation module 610, and pre-processes the multi-source spatial data to obtain a data set that can support theoretical research on city site selection; the spatial division module 620 performs spatial subdivision on a target region based on a division method of a regular polygon to obtain a spatial grid unit, and performs full-coverage and gapless spatial subdivision on the range of the experimental area; the explicit feature extraction module 630 extracts explicit features in each spatial grid unit one by one based on the multi-source spatial data; the implicit feature extraction module 640 constructs a global spatial topology graph, samples the coupled implicit features between the spatial grid units in combination with a VGAE model, constructs an implicit feature matrix, and realizes sampling of the coupled implicit feature information between the spatial grid units and construction of the feature matrix by integrating more comprehensive influence factor analysis; and the prediction module 650 fuses the explicit feature matrix and the implicit feature matrix, and performs site selection prediction of the city emergency shelter based on an RF model, so that the site selection result is more uniform, the coverage is wider, the reliability and accuracy of the site selection result are improved, and the over-concentration or sparse distribution of the emergency shelter can be avoided to a greater extent, thereby improving the overall spatial suitability.

[0199] In one embodiment, the implicit feature extraction module 640 comprises:

[0200] The topology graph construction unit 641 is configured to construct a global spatial topology graph according to the explicit features in the spatial grid unit and the adjacency relationship of the regular polygon.

[0201] The matrix construction unit 642 is configured to construct the explicit features of the city structure, the geographical environment and the social attribute into a feature matrix of the node.

[0202] The low-dimensional space operation unit 643 is configured to learn the explicit features of each spatial grid unit through an encoder containing two layers of graph convolutional neural networks in the VGAE model to obtain the mean and variance of the low-dimensional space representation.

[0203] The implicit feature sampling unit 644 samples the implicit features in the low-dimensional space through a decoder, and reconstructs the spatial topology graph through an inner product operation to construct an implicit feature matrix.

[0204] The embodiment of the application is a device embodiment corresponding to the above-mentioned method embodiment, and the specific operations of each module processing step can be understood with reference to the description of the method embodiment, which will not be repeated here.

[0205] As Figure 9As shown, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the urban emergency shelter site selection method based on the VGAE-RF model in the above-mentioned embodiment, or, when executed by a processor, implements the urban emergency shelter site selection method based on the VGAE-RF model in the above-mentioned embodiment.

[0206] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0207] Each embodiment in this specification is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the device or system embodiments, since they are basically similar to the method embodiments, the description is relatively simple. For the relevant parts, refer to the partial description of the method embodiments. The device and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. A person of ordinary skill in the art can understand and implement it without making any creative efforts.

[0208] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and the contents not described in detail in the specification of the present invention are common knowledge to those skilled in the art.

Claims

1. A method for selecting an urban emergency shelter site based on the VGAE-RF model, characterized in that: The following steps are involved: Access multi-source spatial data from open data sources; The target area is spatially divided based on the regular polygon partitioning method to obtain spatial grid units; Extracting the explicit features of the urban structure, geographical environment, and social attributes within each spatial grid unit one by one based on the multi-source spatial data, and constructing an explicit feature matrix; Constructing a global spatial topology map, combining it with the VGAE model, sampling the implicit features of the coupling between the spatial grid units, and constructing an implicit feature matrix; The explicit feature matrix and the implicit feature matrix are integrated, and the location prediction of urban emergency shelters is performed based on the RF model; The dominant features of the geographical environment include average altitude, slope mean, slope standard deviation, and elevation standard deviation. The method of extracting the dominant features of the geographical environment within each spatial grid cell one by one based on the multi-source spatial data to construct a dominant feature matrix includes the following steps: Performing linear normalization processing on the average altitude, the slope mean, the slope standard deviation, and the elevation standard deviation by column to construct a geographical environment feature matrix; The method of constructing a global spatial topology map and combining it with a VGAE model to sample the implicit features of the coupling between the spatial grid units and construct an implicit feature matrix includes the following steps: Constructing a global spatial topology map based on the dominant features inside the spatial grid cells and the adjacency relationship of the regular polygons; The explicit characteristics of urban structure, geographical environment and social attributes are constructed as a feature matrix of nodes; In the VGAE model, the explicit features of each spatial grid cell are learned through an encoder consisting of a two-layer graph convolutional neural network to obtain the mean and variance of the low-dimensional spatial representation; The latent features are sampled in the low-dimensional space by the decoder, and the spatial topology is reconstructed by the inner product operation method to construct the latent feature matrix.

2. The method for selecting a site for an urban emergency shelter according to claim 1, wherein: The dominant features of the urban structure include road network density features and building features. The method of extracting the dominant features of the urban structure within each spatial grid unit one by one based on the multi-source spatial data and constructing a dominant feature matrix includes the following steps: The road network density is comprehensively calculated by the ratio of the total length of the four types of roads in the spatial grid unit to the area of ​​the spatial grid unit and the road network density is linearly normalized to construct a road network density feature matrix; The building density is determined by the number of buildings per unit area, the average building height is determined by dividing the total building height by the number of buildings, and the building density and the average building height are linearly normalized to construct a building feature matrix.

3. The method for selecting a site for an urban emergency shelter according to claim 1, wherein: The dominant features of the social attributes include population density and POI density. The method of extracting the dominant features of the social attributes within each spatial grid unit one by one based on the multi-source spatial data and constructing a dominant feature matrix includes the following steps: The population density is calculated using the ratio of the total population of a spatial grid cell to the area of ​​the spatial grid cell, and a population density feature matrix is ​​constructed; Construct the POI quantity matrix, calculate the POI density matrix, and construct the POI density feature matrix by normalizing, extracting and reclassifying the POI density matrix.

4. The method for selecting a site for an urban emergency shelter according to claim 1, wherein: The loss function of the VGAE-RF model consists of two parts: a weighted cross entropy loss function and the KL divergence between the latent space distribution and the standard normal distribution.

5. A device for selecting an urban emergency shelter site based on a VGAE-RF model, characterized in that: include: Data generation module, used to obtain multi-source spatial data from open data sources; The spatial partitioning module is used to spatially partition the target area based on the regular polygon partitioning method to obtain spatial grid units; The dominant feature extraction module is used to extract the dominant features of the urban structure, geographical environment, and social attributes of each spatial grid cell one by one based on the multi-source spatial data, and construct a dominant feature matrix. The dominant features of the geographical environment include average altitude, slope mean, slope standard deviation, and elevation standard deviation. The specific configuration is as follows: Performing linear normalization processing on the average altitude, the slope mean, the slope standard deviation, and the elevation standard deviation by column to construct a geographical environment feature matrix; The implicit feature extraction module is used to construct the global spatial topology map. In combination with the VGAE model, it samples the implicit features between the couplings of the spatial grid units and constructs the implicit feature matrix. The specific configuration is as follows: Constructing a global spatial topology map based on the dominant features inside the spatial grid cells and the adjacency relationship of the regular polygons; The explicit characteristics of urban structure, geographical environment and social attributes are constructed as a feature matrix of nodes; In the VGAE model, the explicit features of each spatial grid cell are learned through an encoder consisting of a two-layer graph convolutional neural network to obtain the mean and variance of the low-dimensional spatial representation; Sampling latent features in the low-dimensional space through a decoder, and reconstructing a spatial topology map through an inner product operation method to construct a latent feature matrix; The prediction module is used to fuse the explicit feature matrix and the implicit feature matrix and perform site selection prediction of urban emergency shelters based on the RF model.

6. The urban emergency shelter site selection device according to claim 5, characterized in that: The latent feature extraction module includes: A topology map construction unit, configured to construct a global spatial topology map based on the dominant features within the spatial grid cells and the adjacency relationship of the regular polygons; A matrix construction unit is used to construct the explicit characteristics of urban structure, geographical environment and social attributes into a feature matrix of nodes; The low-dimensional spatial operation unit is used in the VGAE model to learn the explicit features of each spatial grid cell through an encoder containing a two-layer graph convolutional neural network to obtain the mean and variance of the low-dimensional spatial representation; The latent feature sampling unit samples latent features in the low-dimensional space through a decoder, reconstructs a spatial topology map through an inner product operation method, and constructs a latent feature matrix.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method for selecting an urban emergency shelter site based on the VGAE-RF model as described in any one of claims 1 to 4 is implemented.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for selecting an urban emergency shelter site based on a VGAE-RF model as claimed in any one of claims 1 to 4 are implemented.

Images