Evaluation and planning integrated mountain village and town emergency evacuation decision-making method, device, equipment and medium

Abstract

The invention relates to a mountain village and town emergency evacuation decision-making method and device integrating evaluation and planning, equipment and a medium. The method comprises the following steps: acquiring multi-source heterogeneous data of a target mountainous village and town, and carrying out standardized preprocessing on the multi-source heterogeneous data to obtain multi-source standardized raster data; based on the multi-source standardized raster data, constructing a unit graph structure which takes a space unit as a node and takes a space adjacency relation as an edge; inputting the unit graph structure into a preset double-flow feature enhanced graph convolutional network model, and outputting an emergency evacuation capability assessment score of each space unit in the unit graph structure; diagnosing the key restriction features according to the emergency evacuation capability evaluation score to obtain a diagnosis result; and generating an evacuation space planning scheme based on the diagnosis result. The method can solve the problems that a traditional method is weak in data basis, static assessment is rigid, and planning breaks away from actual risks, achieves the deep fusion of assessment and planning links, and improves the scientificity and dynamic response capability of emergency evacuation decisions of mountainous villages and towns.

Description

Technical Field

[0001] This invention belongs to the field of disaster prevention and mitigation technology, and in particular relates to the method, device, equipment and medium for emergency evacuation decision-making in mountainous villages and towns with integrated assessment and planning. Background Technology

[0002] With the development of disaster prevention and mitigation technologies, disaster response in mountainous villages and towns has received increasing attention. Currently, a spatial perspective-based resilience enhancement approach has emerged in this field, characterized by strengthening regional disaster resilience through long-term, systematic planning. This leads to the current mainstream emergency evacuation assessment and planning methods, which rely on static data analysis and overlay analysis using traditional Geographic Information System (GIS).

[0003] Traditional methods for assessing emergency evacuation capabilities in mountainous villages and towns rely primarily on historical experience, manual surveys, and limited monitoring data. The assessment process typically involves overlaying various spatial data (such as topography, roads, and buildings) into layers, combining expert scoring or simple weighted indicator models to generate a macro-level risk distribution map or evacuation plan. This approach aims to identify risk areas holistically and plan evacuation sites and routes.

[0004] However, current traditional methods have significant problems. Mountainous areas have complex natural environments, with diverse and sudden disasters, and traditional assessment models struggle to effectively integrate multi-source, heterogeneous, and dynamically changing monitoring data. More importantly, existing methods cannot precisely characterize the micro-risk differences caused by complex terrain and fragmented features; for example, their ability to identify key evacuation elements such as residents in shaded areas and trails in vegetated areas is severely inadequate. This results in assessments that are often crude and cannot provide reliable support for rapid and accurate evacuation decisions in the event of a sudden disaster. Furthermore, the lag in the revision of emergency plans and the inadequacy of the material and equipment support system further affect the efficiency of emergency response. Summary of the Invention

[0005] Therefore, it is necessary to provide integrated assessment and planning methods, devices, equipment, and media for emergency evacuation decision-making in mountainous villages and towns to address the aforementioned technical issues.

[0006] Firstly, this application provides an integrated assessment and planning method for emergency evacuation decision-making in mountainous villages and towns, including:

[0007] S1. Acquire multi-source heterogeneous data of the target mountainous villages and towns, and perform standardized preprocessing on the multi-source heterogeneous data to obtain multi-source standardized raster data; the multi-source heterogeneous data includes multimodal remote sensing images, digital elevation model data, geographic information vector data, disaster historical data and planning attribute data;

[0008] S2. Based on multi-source standardized raster data, construct a cell graph structure with spatial cells as nodes and spatial adjacency relationships as edges;

[0009] S3. Input the unit graph structure into the preset dual-stream feature-enhanced graph convolutional network model, and output the emergency evacuation capability evaluation score of each spatial unit in the unit graph structure; the dual-stream feature-enhanced graph convolutional network model includes a visual terrain feature branch, a semantic attribute feature branch, a cross-stream attention interaction module, and a multilayer perceptron classifier;

[0010] S4. Based on the emergency evacuation capability assessment score, diagnose key limiting characteristics and obtain diagnostic results;

[0011] S5. Generate an evacuation space planning scheme based on the diagnostic results; the evacuation space planning scheme includes evacuation space planning and dynamic route decision suggestions.

[0012] In one embodiment, based on multi-source standardized raster data, a cell graph structure is constructed with spatial cells as nodes and spatial adjacency relationships as edges, including:

[0013] S11. Perform superpixel segmentation on the optical image in the multimodal remote sensing image, divide it into multiple spatially continuous and feature-similar superpixel regions, define each superpixel region as a spatial unit node, and obtain a node set.

[0014] S12. Extract visual appearance features and terrain features from the multi-source standardized raster data within the geographic space corresponding to each spatial unit node in the node set.

[0015] S13. Overlay and analyze the spatial location and geographic information vector data, disaster history data and planning attribute data of each spatial unit node to extract geographic semantic features, including road density and building density.

[0016] S14. The visual appearance features, terrain features and geographic semantic features corresponding to each spatial unit node are spliced ​​and fused to generate the fused feature vector of each spatial unit node, and the fused feature vectors of all spatial unit nodes are used to form a node feature matrix.

[0017] S15. Based on the spatial adjacency relationship of all spatial unit nodes, establish undirected edges between adjacent nodes to form the edge set of the graph, and construct the unit graph structure based on the node set, edge set, and node feature matrix.

[0018] In one embodiment, the cell graph structure is input into a preset dual-stream feature-enhanced graph convolutional network model, and the emergency evacuation capability evaluation score of each spatial cell in the cell graph structure is output, including:

[0019] S21. Input the node feature matrix in the unit graph structure into the dual-stream feature enhancement graph convolutional network model, and split the fused feature vector of each spatial unit node into visual terrain feature sub-vector and semantic attribute feature sub-vector.

[0020] S22. Input the visual terrain feature sub-vectors into the visual terrain feature branch to obtain the visual enhancement features; the visual terrain feature branch includes the first image convolutional layer;

[0021] S23. Input the semantic attribute feature sub-vectors into the semantic attribute feature branch to obtain the semantic enhancement features; the semantic attribute feature branch includes the second graph convolutional layer;

[0022] S24. Through the cross-stream attention interaction module, calculate the correlation weight between visual enhancement features and semantic enhancement features, and perform weighted fusion of visual enhancement features and semantic enhancement features based on the correlation weight to generate a unified deep feature representation for each spatial unit node.

[0023] S25. Input the unified deep feature representation into the multilayer perceptron classifier to obtain the emergency evacuation capability assessment score of each spatial unit node corresponding to the spatial unit.

[0024] In one embodiment, the two-stream feature enhancement map convolutional network model is trained using the following method:

[0025] S31. Obtain multi-source heterogeneous data from multiple historical mountainous village and town cases to obtain a training graph sample set;

[0026] S32. For each training map sample in the training map sample set, use the comprehensive evaluation method combining the analytic hierarchy process and the entropy weight method to determine the weight of each indicator in the corresponding emergency evacuation capability evaluation index system, and combine the fuzzy comprehensive evaluation method to calculate the emergency evacuation capability score for each spatial unit in the training map sample, which serves as the supervised training label for the training map sample.

[0027] S33. Using the cell graph structure of the training image samples as input, the corresponding supervised training labels as the target output, and the mean squared error as the loss function, the parameters of the two-stream feature enhancement graph convolutional network model are iteratively optimized and trained using the backpropagation algorithm; where the expression for the mean squared error is:

[0028]

[0029] In the formula, Indicates mean square error. This represents the total number of spatial unit nodes in a training batch. Indicates the first Supervised training label scores corresponding to each spatial unit node The model represents the first The emergency evacuation capability assessment score predicted by each spatial unit node;

[0030] S34. When the model's performance on the validation set stabilizes, stop training, save the model parameters, and obtain the final two-stream feature-enhanced graph convolutional network model.

[0031] In one embodiment, based on the emergency evacuation capability assessment score, key limiting characteristics are diagnosed to obtain diagnostic results, including:

[0032] S41. Based on the preset weak threshold, select units from all spatial units whose emergency evacuation capability assessment scores are lower than the weak threshold and mark them as target weak units.

[0033] S42. For each target weak unit, obtain the original fusion feature vector of the corresponding node in the unit graph structure;

[0034] S43. Calculate the gradient value of the contribution of each feature dimension in the fused feature vector to the emergency evacuation capability assessment score;

[0035] S44. Sort the data according to the absolute size of the contribution gradient values, identify the feature dimensions that have the main negative impact on the emergency evacuation capability assessment score, determine the original data features represented by the feature dimensions as key constraint features, and summarize to generate diagnostic results.

[0036] In one embodiment, evacuation space planning is obtained through the following method:

[0037] S51. Analyze the diagnostic results. If the key limiting features involve the accessibility of emergency shelters, locate the geographical location of the relevant vulnerable units. Based on slope data, historical disaster data, and land use data, conduct a multi-criteria suitability analysis in the surrounding buffer zone to generate site selection recommendations.

[0038] S52. Analyze the diagnostic results. When the key constraint features involve the connectivity of the road network, based on the topological connection relationship of the unit graph structure, use network analysis methods to identify the key missing road segments or bottleneck road segments that restrict the connectivity of weak units and generate road planning suggestions.

[0039] S53. Integrate site selection recommendations with road planning recommendations to form an evacuation space plan.

[0040] In one embodiment, dynamic path decision recommendations are obtained through the following method:

[0041] S61. After receiving a specific natural disaster early warning information, adjust the weights or states of relevant features in the multi-source heterogeneous data according to the disaster type, and re-execute S2 to S3 to obtain the updated dynamic evaluation score; the updated dynamic evaluation score is used to characterize the dynamic emergency evacuation capacity evaluation score under the disaster scenario.

[0042] S62. Based on the updated dynamic assessment score, identify the current risk area and obtain the population distribution information and real-time capacity status of emergency shelters within the corresponding area.

[0043] S63. Taking residential areas within the risk zone as the starting point of the route and available emergency shelters as the ending point, the dynamic evaluation score is used as a constraint on the route's traffic resistance. The shortest path search algorithm is then used to calculate and generate recommended evacuation routes.

[0044] S64. Package the recommended evacuation routes with the estimated evacuation time and route capacity assessment information to generate dynamic route decision suggestions.

[0045] Secondly, this application also provides an integrated assessment and planning emergency evacuation decision-making device for mountainous villages and towns, comprising:

[0046] The data acquisition and standard preprocessing module is used to acquire multi-source heterogeneous data of target mountain villages and towns, and to perform standard preprocessing on the multi-source heterogeneous data to obtain multi-source standardized raster data; the multi-source heterogeneous data includes multimodal remote sensing images, digital elevation model data, geographic information vector data, disaster historical data and planning attribute data;

[0047] The cell graph construction module is used to construct a cell graph structure based on multi-source standardized raster data, with spatial cells as nodes and spatial adjacency relationships as edges;

[0048] The evacuation capability assessment module is used to input the unit graph structure into a preset dual-stream feature-enhanced graph convolutional network model and output the emergency evacuation capability assessment score of each spatial unit in the unit graph structure. The dual-stream feature-enhanced graph convolutional network model includes a visual terrain feature branch, a semantic attribute feature branch, a cross-stream attention interaction module, and a multilayer perceptron classifier.

[0049] The constraint diagnosis module is used to diagnose key constraint characteristics based on the emergency evacuation capability assessment score and obtain diagnostic results.

[0050] The evacuation space planning scheme generation module is used to generate evacuation space planning schemes based on diagnostic results; the evacuation space planning schemes include evacuation space planning and dynamic route decision suggestions.

[0051] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0052] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0053] The aforementioned integrated assessment and planning method, device, equipment, and media for emergency evacuation decision-making in mountainous villages and towns integrate multi-source heterogeneous information such as multimodal remote sensing imagery, digital elevation models, geographic information vector data, disaster history, and planning attributes. It constructs a graph structure with spatial units as nodes and adjacency relationships as edges, and utilizes a dual-stream feature-enhanced graph convolutional network model to conduct refined assessments of the emergency evacuation capabilities of mountainous villages and towns. This effectively overcomes the limitations of traditional methods, such as relying on a single data source and having an incomplete assessment foundation. This method can diagnose key constraints, providing a scientific basis for evacuation spatial planning, such as optimizing the selection of refuge sites and road network connectivity. Simultaneously, by combining real-time disaster early warning with dynamic adjustments to the assessment results, it generates adaptive evacuation route suggestions, addressing the shortcomings of traditional methods that are static, rigid, and unable to cope with sudden changes. This improves the accuracy, dynamism, and practicality of emergency decision-making, providing efficient and reliable technical support for disaster prevention and mitigation in mountainous villages and towns. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart illustrating an emergency evacuation decision-making method for mountainous villages and towns that integrates assessment and planning in one embodiment.

[0056] Figure 2 This is a schematic diagram of the structure of an evacuation capacity assessment system for mountainous villages and towns in one embodiment;

[0057] Figure 3 This is a flowchart illustrating a method for assessing evacuation capacity in mountainous villages and towns, as described in one embodiment.

[0058] Figure 4 This is a schematic diagram of the structure of an emergency evacuation decision-making device for mountainous villages and towns that integrates assessment and planning in one embodiment. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0060] refer to Figure 1 The application presents a flowchart illustrating the integrated assessment and planning method for emergency evacuation decision-making in mountainous villages and towns, which includes the following steps:

[0061] S1. Obtain multi-source heterogeneous data of the target mountainous villages and towns, and perform standardized preprocessing on the multi-source heterogeneous data to obtain multi-source standardized raster data.

[0062] Optionally, multi-source heterogeneous data includes multimodal remote sensing imagery, digital elevation model data, geographic information vector data, historical disaster data, planning attribute data, and risk source data.

[0063] Optionally, when acquiring multi-source heterogeneous data of target mountain villages and towns, it is necessary to combine the core needs of emergency evacuation for mountain disasters and collect targeted data covering all dimensions, including terrain, land features, disaster history, and planning layout. Among them, high-resolution satellite imagery and low-altitude drone aerial imagery are prioritized for multimodal remote sensing imagery. Satellite imagery is used to acquire macroscopic information such as topography, vegetation cover, and building distribution over a wide area, while drone imagery focuses on key microscopic areas such as densely populated areas, narrow mountain roads, and steep slopes, compensating for the shortcomings of satellite imagery in terms of shadow occlusion and detail depiction. Digital Elevation Model (DEM) data uses DEM data of appropriate resolution to ensure accurate extraction of core topographic parameters such as slope, aspect, and elevation difference. Geographic information vector data is obtained from local natural resources departments and includes vector layers of roads, water systems, settlements, public facilities, and medical points, supplemented by data on the distribution of infrastructure such as power facilities and communication base stations. Historical disaster data is retrieved from emergency management departments and meteorological departments, covering detailed records of the occurrence time, location, impact range, and degree of damage of various disasters in and around the target area. Risk source data includes the location and safe distance of fixed risk sources such as mines, tailings ponds, chemical plants, gas stations, firework and firecracker warehouses, and dams, as well as mobile risk sources such as hazardous materials transportation routes. Accident information is retrieved from the monitoring and alarm systems of departments such as emergency management, ecology and environment, industry and information technology, and transportation (for production safety accidents). Meteorological data is obtained from the meteorological department. Planning attribute data includes official documents such as local village and town master plans, disaster prevention and mitigation special plans, and shelter layout plans, from which key information such as shelter levels, evacuation route plans, and material reserve location are extracted.

[0064] Data standardization preprocessing requires a step-by-step approach to eliminate format differences, spatial reference differences, and accuracy differences among multi-source data to ensure data fusion compatibility. First, data format unification is performed by converting vector data to raster format using the feature-to-raster tool in professional geographic information software. Remote sensing imagery and DEM data are converted to a unified, universal raster format using professional remote sensing processing software. Second, coordinate system unification is implemented. Based on the administrative area of ​​the target mountainous villages and towns, a suitable local projection coordinate system is selected. Geographic registration tools are used to select evenly distributed ground control points to achieve spatial alignment of all data, ensuring that registration errors are controlled within a reasonable range. Next, data cleaning is performed, using the 3σ criterion to remove noise points from remote sensing imagery and abnormal elevation values ​​from DEM data. Missing data is supplemented using Kriging interpolation or nearest neighbor interpolation. Finally, normalization is performed, using the min-max normalization method to convert attribute data of different dimensions to the [0,1] interval. This normalization process ensures balanced feature weights during subsequent model training. After the above preprocessing, multi-source standardized raster data is obtained. The spatial resolution, coordinate system, and numerical range of each data layer are completely consistent and can be directly used for subsequent graph structure construction.

[0065] S2. Based on multi-source standardized raster data, construct a cell graph structure with spatial cells as nodes and spatial adjacency relationships as edges.

[0066] Optionally, when constructing a unit graph structure based on multi-source standardized raster data, the core is to achieve refined discretization and topological relationship modeling of mountainous village and town spaces, ensuring that the graph structure can accurately reflect the attribute characteristics and spatial relationships of spatial units. First, spatial units are divided, using preprocessed raster data as basic spatial units. Each raster corresponds to an initial node. For areas with continuous terrain and similar attributes, adjacent raster data can be merged using a region growing algorithm. The merging condition is that the slope difference and building density difference between adjacent raster data are within a set threshold, avoiding distortion of micro-risk characterization due to excessively large units. For areas with complex terrain, the original raster data is retained as independent spatial units to ensure accurate capture of the impact of local terrain on evacuation.

[0067] Each spatial unit serves as a node in the graph structure, and its feature vector is composed of the attribute values ​​of the multi-source standardized raster data of the corresponding unit. The specific dimensions include terrain attributes, land feature attributes, disaster attributes, planning attributes, and risk source attributes. Each attribute value is obtained by extracting the mean or maximum value of the raster data within the corresponding spatial unit, ultimately forming a node feature vector with fixed dimensions.

[0068] Spatial adjacency relationships, acting as edges in the graph structure, are determined by whether two spatial units are physically adjacent. An 8-neighborhood rule is used: if two units share a boundary or a vertex, they are considered adjacent and a connecting edge is constructed. For adjacent units separated by insurmountable obstacles such as rivers or cliffs, their connecting edges are removed. Obstacle information is determined from the water system and terrain slope layers in the geographic information vector data. Edge weights are determined by calculating the feature similarity between two adjacent units using the cosine similarity formula, as follows:

[0069]

[0070] in, Represents a node With nodes Feature similarity between Represents a node eigenvectors, Represents a node eigenvectors, This represents the dot product of two eigenvectors. Represents a node The L2 norm of the eigenvectors Represents a node The L2 norm of the eigenvectors. Higher similarity results in larger edge weights. These weights reflect the ease of passage between adjacent units; units with high similarity have low passage resistance and weights approaching 1, while units with low similarity have high passage resistance and weights approaching 0. Through these steps, a unit graph structure containing several nodes and edges is finally constructed. This structure preserves the spatial topological relationships of mountain villages and integrates multi-dimensional attribute features, providing a foundation for subsequent feature enhancement and evaluation.

[0071] S3. Input the unit graph structure into the preset dual-stream feature enhancement graph convolutional network model, and output the emergency evacuation capability evaluation score of each spatial unit in the unit graph structure.

[0072] Optionally, the dual-stream feature enhancement graph convolutional network model includes a visual terrain feature branch, a semantic attribute feature branch, a cross-stream attention interaction module, and a multilayer perceptron classifier.

[0073] Optionally, the preset dual-stream feature enhancement map convolutional network model adopts a branch extraction-cross-stream fusion-classification output architecture, which is specifically designed for multi-source feature fusion and evacuation capacity assessment of spatial units in mountainous villages and towns. The model has been trained to convergence using previously labeled sample data of evacuation capacity in mountainous villages and towns. During the training process, the cross-entropy loss function is used to optimize the model parameters to ensure that the assessment accuracy meets the needs of practical applications.

[0074] The visual terrain feature branch is mainly responsible for extracting visual terrain features from multi-source data. The input consists of remote sensing image slices corresponding to nodes in the unit graph structure and terrain rendering maps derived from DEM. The branch adopts a multi-layer convolutional neural network structure. The first layer uses convolutional kernels to extract basic texture features and ReLU activation function. The second layer uses convolutional kernels to extract mid-level terrain features, followed by batch normalization layers and max pooling layers. Subsequent layers adopt a residual block structure, with each residual block containing two convolutional layers. Skip connections are used to solve the gradient vanishing problem, and the final output is a fixed-dimensional visual terrain feature vector.

[0075] The semantic attribute feature branch is responsible for extracting structured semantic attribute features. The input is the multi-dimensional attribute feature vector of the node in the unit graph structure. The branch adopts a multi-layer fully connected network. The first layer maps the input feature vector to a fixed dimension and uses LeakyReLU as the activation function to avoid the problem that the gradient of ReLU is 0 in the negative interval. The second layer maps to a fixed dimension and is followed by a Dropout layer to prevent overfitting. The third layer maintains the fixed dimension output to obtain the semantic attribute feature vector.

[0076] The cross-stream attention interaction module is used to achieve adaptive fusion of features from two branches. Its core is to calculate the association weights of the two feature vectors through an attention mechanism, highlighting key feature information for assessing dispersal capabilities. Specifically, the process involves first inputting the visual terrain feature vector and semantic attribute feature vector into the attention computation layer, and obtaining the feature association matrix through dot product operations; then, normalizing the association matrix using the softmax function to obtain the attention weight matrix; next, weighted summing of the two feature vectors with the attention weight matrix to obtain the enhanced visual terrain features and semantic attribute features; finally, a concatenation operation merges the two enhanced feature vectors into a fixed-dimensional fused feature vector.

[0077] A multilayer perceptron classifier receives fused feature vectors and outputs an evaluation score for the emergency evacuation capability of each spatial unit. The classifier consists of two fully connected layers and one output layer. The first layer maps the fused features to a fixed dimension using ReLU activation. The second layer maps to a fixed dimension followed by a batch normalization layer. The output layer uses a linear activation function to output an evaluation score within a set range; a higher score indicates stronger evacuation capability. After inputting the constructed unit graph structure into the trained model, the model updates node features based on node features and edge weights through graph convolution operations to achieve accurate evaluation of the evacuation capability of each spatial unit, ultimately outputting an evaluation score for each spatial unit.

[0078] S4. Based on the emergency evacuation capability assessment score, diagnose key limiting characteristics and obtain diagnostic results.

[0079] Optionally, when diagnosing key constraints based on emergency evacuation capacity assessment scores, a diagnostic logic of score grading, feature association, and constraint localization needs to be established to ensure that the diagnostic results accurately reflect the core issues affecting evacuation capacity. First, assessment scores are graded. Based on the actual needs of emergency evacuation in mountainous villages and towns, the assessment scores are divided into five levels. Spatial units with insufficient evacuation capacity are then selected as key diagnostic targets through this grading process.

[0080] Subsequently, a correlation model between evaluation scores and multi-source features was established. A random forest algorithm was used to analyze the influence weight of each input feature on the evaluation score. Specifically, the feature vectors of all spatial units were used as input, and the evaluation score as the output to train the random forest model. After training, the importance weight of each feature was calculated using the Gini coefficient; a larger weight value indicates a greater impact of that feature on the evacuation capability assessment. Simultaneously, for key diagnostic targets, a local interpretability model was used to analyze the feature contribution of each unit, i.e., the positive or negative influence of each feature value on the unit's evaluation score, thereby identifying potential limiting factors.

[0081] The diagnosis of key constraints requires combining the importance weight of the characteristics with the contribution of local characteristics to identify the core constraint types of units with insufficient evacuation capacity at different levels. Specific classifications include terrain constraints, transportation constraints, infrastructure constraints, environmental constraints, and planning deficiencies. For each key diagnostic target, the specific manifestations of the constraints need to be accurately located, and the core attributes and influencing dimensions of the constraints need to be clarified.

[0082] The diagnostic results are presented in a structured format, with each key diagnostic target corresponding to a diagnostic record. This record includes the spatial unit number, assessment score and level, core constraint type, specific constraint parameter, degree of influence of the constraint, and scope of influence. Through batch diagnostics, a spatial distribution map of the constraints on the evacuation capacity of target mountain villages and towns is generated, clarifying the concentrated area of ​​weak points and the types of core constraints, providing precise targeting for subsequent planning schemes.

[0083] S5. Generate an evacuation space planning scheme based on the diagnostic results.

[0084] Optionally, the evacuation space planning scheme includes evacuation space planning and dynamic route decision recommendations.

[0085] Optionally, when generating evacuation space planning schemes based on diagnostic results, the principles of targeted governance, adapting to local conditions, and safety and efficiency should be followed. Differentiated planning measures should be formulated for different types of constraints, while ensuring that the planning schemes are compatible with the mountainous terrain and existing infrastructure and are feasible.

[0086] Evacuation space planning mainly includes three core parts: optimized layout of refuge sites, upgraded evacuation routes, and supplementary emergency facilities. For optimized refuge site layout, new refuge sites are planned for areas with weak infrastructure, taking into account requirements such as terrain flatness, accessibility, and site area. The locations of new sites are determined using the minimum service distance model of GIS spatial analysis tools to ensure coverage of all areas with weak evacuation capacity. For existing refuge sites, their carrying capacity is assessed based on diagnostic results, and sites with insufficient capacity are expanded and upgraded, with the addition of necessary emergency infrastructure. Simultaneously, emergency material reserve points are planned around refuge sites, with the types of reserve materials determined based on the number of people served to ensure rapid deployment.

[0087] For upgrading evacuation routes, categorized improvement measures will be implemented for vulnerable areas constrained by traffic or terrain: Existing narrow roads will be widened and upgraded using appropriate hardening methods, adjusting road slopes and turning radii to meet emergency passage needs; in areas lacking evacuation routes, new pedestrian evacuation paths will be planned based on terrain conditions, using anti-slip paving materials and installing protective facilities on steep slopes; obstructions to evacuation routes will be cleared to ensure unobstructed visibility; emergency signs will be installed at key points along evacuation routes, displaying necessary evacuation information. Furthermore, for environmentally constrained areas, hazardous areas around evacuation routes will be cleared, reinforced with appropriate support measures, and drainage and diversion facilities will be installed in disaster-prone areas to reduce disaster risk.

[0088] Dynamic route decision-making recommendations are based on diagnostic results and real-time disaster monitoring data, constructing a dynamic route planning model to achieve adaptive optimization of evacuation routes. Model inputs include the spatial distribution of evacuation capacity assessment scores obtained from the diagnosis, real-time disaster monitoring data, and real-time traffic conditions. The route planning algorithm employs an improved A* algorithm, using the evacuation capacity assessment score as the route weight and introducing a disaster risk penalty coefficient. The algorithm aims to calculate the optimal evacuation route from each household to the nearest refuge. The specific calculation process involves first discretizing the target area into a grid map, with the weight of each grid determined by the evacuation capacity score and the disaster risk penalty coefficient; then, starting from the household location and ending at the refuge, the optimal route is searched using the A* algorithm, avoiding areas with extremely poor evacuation capacity and areas affected by real-time disasters during the route search; for areas with concentrated multiple households, shared evacuation routes are planned to avoid localized congestion.

[0089] Dynamic route decision-making recommendations also include evacuation sequence planning. Based on the evacuation capacity assessment results, evacuation areas are divided into priority evacuation zones, secondary priority evacuation zones, and routine evacuation zones. Priority evacuation zones must initiate evacuation immediately after the disaster warning is issued. Secondary priority evacuation zones are initiated after the priority evacuation zones are completed. Routine evacuation zones are initiated flexibly according to the development of the disaster. At the same time, for special groups, dedicated escort evacuation routes are planned, and the responsible escort personnel, escort routes, and assembly points are clearly defined.

[0090] The final evacuation space planning scheme comprises three parts: a planning document, a spatial layout map, and a dynamic route navigation map. The planning document details the planning objectives, implementation scope, specific measures, construction timeline, investment estimates, and responsible parties. The spatial layout map, drawn using GIS, marks the location and extent of newly added or renovated shelters, evacuation routes, and emergency supply points. The dynamic route navigation map is an interactive electronic map that updates evacuation routes based on real-time disaster data and supports simultaneous viewing on multiple devices. The scheme also includes supporting implementation measures, including financial, technical, and organizational guarantees, to ensure the plan's effective implementation and comprehensively enhance the emergency evacuation capabilities of the target mountainous villages and towns.

[0091] The aforementioned integrated assessment and planning method for emergency evacuation decision-making in mountainous villages and towns achieves accurate assessment of emergency evacuation capabilities through standardized processing of multi-source heterogeneous data and construction of superpixel unit graphs, combined with dual-stream feature-enhanced graph convolutional networks. It targets and locates key constraint features based on gradient contribution analysis, and generates targeted evacuation space planning that coordinates refuge sites and road networks. Furthermore, it can dynamically adjust parameters based on disaster early warnings to output evacuation route suggestions adapted to real-time risks. This effectively solves problems such as the crudeness of traditional emergency evacuation assessments in mountainous villages and towns, the ambiguity in the diagnosis of constraint factors, the insufficient targeting of planning schemes, and the lag in route decision-making under disaster scenarios. It improves the scientific nature and timeliness of emergency evacuation system construction, providing reliable technical support for disaster prevention, mitigation, and emergency management in mountainous villages and towns.

[0092] To further illustrate the solutions of this application, a specific embodiment applied to the above-mentioned integrated assessment and planning method for emergency evacuation decision-making in mountainous villages and towns will be described below, such as... Figure 2 The diagram shows a structural schematic of an evacuation capacity assessment system 10 for mountainous villages and towns, including:

[0093] 1. User registration module 11, used for new user registration;

[0094] 2. User login module 12, used for user login;

[0095] 3. Archives Management Module 13 is used to collect historical disaster data, geographic information data, disaster prevention and mitigation planning data, and risk source data for mountainous villages and towns;

[0096] The aforementioned historical disaster data includes information on historical collapses, landslides, mudslides, floods, forest fires, earthquakes, etc.

[0097] The aforementioned geographic information data includes GIS maps, slope, vegetation coverage, buildings, roads, rivers, schools, and other information.

[0098] The aforementioned disaster prevention and mitigation planning data includes information on shelters, infrastructure, medical institutions, and population.

[0099] The aforementioned risk source data includes the location and danger radius of fixed risk sources such as mines, tailings ponds, chemical plants, gas stations, firework and firecracker warehouses, and dams, as well as mobile risk sources such as major traffic arteries, especially hazardous materials transportation routes.

[0100] 4. Emergency evacuation capacity evaluation module 14 is used to evaluate based on historical disaster data of mountainous villages and towns to obtain evaluation results;

[0101] The above evaluation includes establishing a GCNs model, assessing emergency evacuation capabilities, and visualizing the distribution of emergency evacuation capabilities within the study area.

[0102] 5. Emergency evacuation space planning module 15, used to develop targeted emergency plans based on evaluation results;

[0103] The system also includes an evaluation index system unit, a comprehensive evaluation scoring unit, and an emergency shelter planning unit. The evaluation index system unit is used to determine the weights of each indicator factor and the scores of each level using the ANP-EWM (analytic network process, ANP; entropy weight method, EWM) comprehensive evaluation method. The comprehensive evaluation scoring unit includes using the fuzzy comprehensive evaluation method to conduct a comprehensive evaluation and scoring of the emergency evacuation capacity of the study area. The emergency shelter planning unit includes proposing improvement schemes for evacuation space planning in areas with weak emergency evacuation capacity based on the evaluation results.

[0104] To further illustrate the solutions of the embodiments of this application, the following description uses a specific embodiment applied to the above-mentioned evacuation capacity assessment system 10 for mountainous villages and towns. Figure 3 The diagram shows a flowchart of a method for assessing evacuation capacity in mountainous villages and towns, including the following steps:

[0105] S100: Processes remote sensing images, imports them into the system, uses graph convolutional neural network (GCNs) models to evaluate emergency evacuation capabilities, and outputs visualizations.

[0106] S200: Establish an emergency evacuation capability index evaluation system and use the ANP-EWM method to determine the weight of each index factor and the score of each level.

[0107] S300: Use the fuzzy comprehensive evaluation method to comprehensively evaluate and score the emergency evacuation capacity of the study area, and propose improvement schemes for evacuation space planning for areas with weak emergency evacuation capacity.

[0108] The above-mentioned method for assessing the evacuation capacity of mountainous villages and towns establishes a graph convolutional network (GCNs) model to evaluate the emergency evacuation capacity of mountainous villages and towns. It also uses ANP-EWM analysis to analyze the weights of emergency evacuation capacity assessment indicators and fuzzy comprehensive evaluation method to obtain evaluation scores. The method provides suggestions for improving evacuation space planning for areas with weak emergency evacuation capacity, provides a reference for optimizing the evacuation space layout of mountainous villages and towns, improves evacuation efficiency, and enhances the disaster prevention, mitigation and relief capabilities of mountainous villages and towns.

[0109] In an optional embodiment, based on multi-source standardized raster data, a cell graph structure is constructed with spatial cells as nodes and spatial adjacency relationships as edges, including:

[0110] S11. Perform superpixel segmentation on the optical image in the multimodal remote sensing image, dividing it into multiple spatially continuous and feature-similar superpixel regions. Define each superpixel region as a spatial unit node to obtain a node set.

[0111] Optionally, the SLIC superpixel segmentation algorithm is employed. By aggregating pixels with similar color and spatial location features, spatial units that fit geographic entities are formed, avoiding the boundary fragmentation problem of traditional fixed grids. First, the optical image is converted from the RGB color space to the CIELAB space to improve the accuracy of color similarity judgment. After setting superpixel size and compactness parameters, cluster centers are uniformly generated and iteratively optimized. The distance from a pixel to the cluster center is calculated using a formula that combines color and spatial distance.

[0112]

[0113] in For the overall distance, The Euclidean distance in CIELAB space. For spatial Euclidean distance, Preset size for superpixels These are the normalization coefficients. After segmentation, each superpixel region is numbered to generate a unique node ID, forming a node set.

[0114] S12. Extract visual appearance features and terrain features from the multi-source standardized raster data within the geographic space corresponding to each spatial unit node in the node set.

[0115] Optionally, visual appearance features are extracted from preprocessed multimodal remote sensing images, including texture features based on the gray-level co-occurrence matrix, CIELAB spatial color statistical features, and normalized vegetation index features; topographic features are extracted from digital elevation model standardized raster data, including slope, aspect, topographic relief, and surface roughness. After all features are normalized by min-max to eliminate dimensional differences, visual appearance feature vectors and topographic feature vectors for each spatial unit node are formed.

[0116] S13. Overlay and analyze the spatial location and geographic information vector data, disaster history data, and planning attribute data of each spatial unit node to extract geographic semantic features, including road density and building density.

[0117] Optionally, the spatial unit node boundary vectors and various attribute data are first integrated into a coordinate system. GIS intersection analysis is then used to determine the road segments and building areas contained within the nodes, and core geographic semantic features are calculated. Road density is the ratio of the total road length within a node to the node area, expressed by the formula:

[0118]

[0119] in For road density, Let A be the total length of roads within the node, and A be the node area; the building density is the ratio of the building footprint to the node area, calculated using the formula:

[0120]

[0121] in For building density, This represents the total building area within the node. Simultaneously, by combining historical disaster data and planning attribute data, disaster impact characteristics and public service facility characteristics are extracted, and after quantification and coding, a geographic semantic feature vector is formed.

[0122] S14. The visual appearance features, terrain features and geographic semantic features corresponding to each spatial unit node are spliced ​​and fused to generate the fused feature vector of each spatial unit node, and the fused feature vectors of all spatial unit nodes are used to form a node feature matrix.

[0123] Optionally, after verifying the dimensionality of the visual appearance, terrain, and geographic semantic feature vectors, they are concatenated in the order of visual appearance-terrain-geographic semantics to form a fused feature vector. Dimensionality reduction is achieved using principal component analysis to eliminate redundant information. The dimensionality-reduced feature vectors are then normalized and arranged in order of node ID to form a node feature matrix. The number of rows in the matrix represents the number of nodes, and the number of columns represents the dimension of the dimensionality-reduced feature vector.

[0124] S15. Based on the spatial adjacency relationship of all spatial unit nodes, establish undirected edges between adjacent nodes to form the edge set of the graph, and construct the unit graph structure based on the node set, edge set, and node feature matrix.

[0125] Optionally, adjacent nodes are determined through GIS topological relationships. Node pairs with a common edge / point at their boundaries or with a spatial distance less than a preset threshold and no obstructions are considered adjacent, and an undirected edge is established for them. The edge weight is calculated by combining feature similarity and spatial distance, using the following formula:

[0126]

[0127] in, For edge weights, For balance coefficient, For feature cosine similarity, Let be the Euclidean distance between the node centers. All undirected edges form an edge set, which is converted into an adjacency matrix and then normalized using the following formula:

[0128]

[0129] in, For the normalized adjacency matrix, It is a degree matrix and Adjacency matrix No. The row elements are summed. Finally, by combining the node set, the normalized adjacency matrix, and the node feature matrix, the cell graph structure is completed.

[0130] In an optional embodiment, the cell graph structure is input into a preset dual-stream feature-enhanced graph convolutional network model, and the emergency evacuation capability evaluation score of each spatial cell in the cell graph structure is output, including:

[0131] S21. Input the node feature matrix in the unit graph structure into the dual-stream feature enhancement graph convolutional network model, and split the fused feature vector of each spatial unit node into visual terrain feature sub-vectors and semantic attribute feature sub-vectors.

[0132] Optionally, after the node feature matrix is ​​input into the model, sub-vector splitting is performed based on the original concatenation order of the fused feature vector. Since the fused feature vector is concatenated in the order of visual appearance features - terrain features - geographic semantic features, the first two parts are merged into a visual terrain feature sub-vector during splitting, corresponding to the natural geographic appearance and terrain attributes of the spatial unit; the last part is used as a separate semantic attribute feature sub-vector, corresponding to the human geographic attributes of the spatial unit. The splitting process is implemented through dimension indexing, ensuring that the sum of the dimensions of the two sub-vectors is consistent with the dimension of the original fused feature vector, and that no feature information is omitted or redundant, laying the foundation for subsequent dual-branch parallel feature extraction.

[0133] S22. Input the visual terrain feature sub-vector into the visual terrain feature branch to obtain the visual enhancement feature.

[0134] Optionally, the visual terrain feature branch includes a first graph convolutional layer.

[0135] Optionally, the visual terrain feature branch takes the visual terrain feature sub-vector as input and uses the first graph convolutional layer to achieve in-depth feature extraction and neighborhood aggregation. The first graph convolutional layer leverages the adjacency relationships of the unit graph structure to weightedly aggregate the features of each spatial unit node with the features of its neighboring nodes, strengthening the supplementary role of spatial correlation information in visual terrain features. The graph convolution calculation uses the formula...

[0136]

[0137] in, For nodes Visual terrain feature sub-vectors, For the normalized adjacency matrix, The weights of the convolutional kernels in the first convolutional layer are... For the corresponding bias term, The LeakyReLU activation function is used. For nodes The set of neighboring nodes. The output of this calculation. This is a visually enhanced feature, which integrates the visual terrain information of the node itself and its neighbors, thereby improving the feature's representational ability.

[0138] S23. Input the semantic attribute feature sub-vector into the semantic attribute feature branch to obtain the semantic enhancement feature.

[0139] Optionally, the semantic attribute feature branch includes a second graph convolutional layer.

[0140] Optionally, the semantic attribute feature branch and the visual terrain feature branch adopt a parallel structure. The input is a semantic attribute feature sub-vector, and the semantic information is further extracted and the neighborhood is aggregated through the second graph convolutional layer. The core logic of the second graph convolutional layer is consistent with that of the first graph convolutional layer, both of which realize feature aggregation by means of the spatial adjacency relationship of the graph structure. Its calculation formula is as follows:

[0141]

[0142] in, For nodes semantic attribute feature sub-vectors, The kernel weights of the convolutional layer in the second graph are... This corresponds to the bias term; the meanings of the remaining parameters are the same as those in the convolutional layer of the first image. The output... This is a semantic enhancement feature, which can accurately characterize the humanistic semantic attributes and neighborhood association features of spatial units.

[0143] S24. Through the cross-stream attention interaction module, calculate the correlation weight between visual enhancement features and semantic enhancement features, and perform weighted fusion of visual enhancement features and semantic enhancement features based on the correlation weight to generate a unified deep feature representation for each spatial unit node.

[0144] Optionally, the cross-stream attention interaction module first calculates the similarity between visual enhancement features and semantic enhancement features through dot product operations to obtain their relevance weights, as shown in the formula:

[0145]

[0146] in, For relevance weights, The feature dimension is used to scale similarity values ​​to avoid numerical overflow. Subsequently, the two enhanced features are weighted and fused based on relevance weights, using the following fusion formula:

[0147]

[0148] in, To unify the deep feature representation, this fusion method integrates two features based on their complementary correlation strength, enhancing information valuable for evacuation capacity assessment while reducing redundant information.

[0149] S25. Input the unified deep feature representation into the multilayer perceptron classifier to obtain the emergency evacuation capability assessment score of each spatial unit node corresponding to the spatial unit.

[0150] Optionally, after the unified deep feature representation is input into the multilayer perceptron classifier, it first undergoes feature mapping and nonlinear transformation through two hidden layers. The ReLU activation function is used in the hidden layers to enhance the nonlinear fitting ability of the model. Finally, the features are mapped to the [0,1] interval through the Sigmoid activation function of the output layer, and then transformed into an emergency evacuation capability evaluation score of 0-100 points through linear transformation, so as to achieve accurate quantification of the evacuation capability of each spatial unit.

[0151] In an optional embodiment, the two-stream feature enhancement map convolutional network model is trained using the following method:

[0152] S31. Obtain multi-source heterogeneous data from multiple historical mountainous village and town cases to obtain a training graph sample set.

[0153] Optionally, historical mountainous village and town cases with different geographical environments, disaster types, and village / town sizes can be selected to ensure the diversity and representativeness of the samples. Multi-source heterogeneous data corresponding to each case are collected, and each case data is converted into a unit graph structure according to the standardized preprocessing process and unit graph construction method described above. The unit graph structures of all cases together constitute the training graph sample set, providing sufficient sample support for model training.

[0154] S32. For each training map sample in the training map sample set, use the comprehensive evaluation method combining the analytic hierarchy process and the entropy weight method to determine the weight of each indicator in the corresponding emergency evacuation capability evaluation index system, and combine the fuzzy comprehensive evaluation method to calculate the emergency evacuation capability score for each spatial unit in the training map sample, which serves as the supervised training label for the training map sample.

[0155] Optionally, an emergency evacuation capacity assessment index system is first constructed, covering core dimensions such as terrain conditions, transportation facilities, public services, and disaster risks. The weights of each index are determined using the analytic hierarchy process (AHP) combined with the entropy weight method. The AHP constructs a judgment matrix through expert scoring to obtain subjective weights, while the entropy weight method obtains objective weights based on the information entropy of the index data. The two are then integrated according to a preset ratio to obtain the comprehensive weight. Combined with the fuzzy comprehensive evaluation method, the qualitative and quantitative evaluations of each index are converted into fuzzy membership degrees by constructing a fuzzy evaluation matrix. Finally, the comprehensive weight is used to calculate the emergency evacuation capacity score for each spatial unit, which serves as the supervised training label for the corresponding training map sample.

[0156] S33. Using the cell graph structure of the training image samples as input, the corresponding supervised training labels as the target output, and the mean squared error as the loss function, the parameters of the two-stream feature enhancement graph convolutional network model are iteratively optimized and trained using the backpropagation algorithm; where the expression for the mean squared error is:

[0157]

[0158] In the formula, Indicates mean square error. This represents the total number of spatial unit nodes in a training batch. Indicates the first Supervised training label scores corresponding to each spatial unit node The model represents the first The emergency evacuation capability assessment score predicted by each spatial unit node.

[0159] In the above expression, Indicates mean square error. This represents the total number of spatial unit nodes in a training batch. Indicates the first Supervised training label scores corresponding to each spatial unit node The model represents the first The emergency evacuation capability assessment score is predicted and output by each spatial unit node. During training, the training graph sample set is divided into a training set and a validation set according to a certain ratio. The gradient of the loss function is backpropagated to each layer of the model through the backpropagation algorithm, and the weights, bias terms, and parameters of the graph convolutional layer and the multilayer perceptron are iteratively updated to optimize the model parameters.

[0160] S34. When the model's performance on the validation set stabilizes, stop training, save the model parameters, and obtain the final two-stream feature-enhanced graph convolutional network model.

[0161] Optionally, the loss value of the validation set is used as the basis for judging model performance. A performance stability threshold is set. When the loss value of the validation set does not decrease significantly for a preset number of consecutive rounds and the fluctuation range is less than the threshold, the model performance is judged to be stable, and iterative training is stopped. The model parameters at this time (including the weights of the two-branch graph convolutional layer, the parameters of the attention module, the parameters of the multilayer perceptron, etc.) are saved to obtain the final dual-flow feature-enhanced graph convolutional network model that can be used for emergency evacuation capability assessment.

[0162] In an optional embodiment, based on the emergency evacuation capability assessment score, key limiting characteristics are diagnosed to obtain diagnostic results, including:

[0163] S41. Based on the preset weak threshold, select units from all spatial units whose emergency evacuation capability assessment scores are lower than the weak threshold and mark them as target weak units.

[0164] Optionally, the weakness threshold is determined by combining relevant standards for emergency evacuation in mountainous villages and towns with actual safety needs, ensuring that the selected areas are those with significant shortcomings in evacuation capacity. By comparing the emergency evacuation capacity assessment scores of each spatial unit with the weakness threshold, spatial units with scores below the threshold are directly marked as target weak units, forming a set of target weak units, which helps to narrow down the analysis scope for subsequent precise diagnosis of limiting characteristics.

[0165] S42. For each target weak unit, obtain the original fusion feature vector of the corresponding node in the unit graph structure.

[0166] Optionally, based on the identifier of the target weak unit, its corresponding node ID in the unit graph structure is matched, and the corresponding original fused feature vector is retrieved from the preset feature storage module through the node ID. The original fused feature vector retains complete feature information of visual appearance, terrain and geographic semantics, without undergoing subsequent processing such as dimensionality reduction, ensuring the accurate correspondence between feature dimensions and original data, laying the foundation for accurate calculation of feature contribution.

[0167] S43. Calculate the gradient value of the contribution of each feature dimension in the fused feature vector to the emergency evacuation capability assessment score.

[0168] Optionally, the contribution gradient value is obtained by backpropagation of the gradient on the trained dual-stream feature enhancement map convolutional network model. The core is to quantify the degree of influence of changes in each feature dimension on the evaluation score. Taking the original fused feature vector of the node corresponding to the weak unit of the target as input, the partial derivative of the emergency evacuation capability evaluation score output by the model with respect to each feature dimension is calculated. This partial derivative is the contribution gradient value of the corresponding feature dimension. The positive or negative sign of the contribution gradient value represents the direction of the feature's influence on the evaluation score, and the absolute value represents the intensity of the influence.

[0169] S44. Sort the data according to the absolute size of the contribution gradient values, identify the feature dimensions that have the main negative impact on the emergency evacuation capability assessment score, determine the original data features represented by the feature dimensions as key constraint features, and summarize to generate diagnostic results.

[0170] Optionally, the contribution gradient values ​​of all feature dimensions are first sorted in descending order of absolute value, prioritizing feature dimensions with a strong impact. Feature dimensions with negative contribution gradient values ​​are then selected, as changes in these dimensions lead to a decrease in the emergency evacuation capacity assessment score, thus negatively constraining evacuation capacity. The original data features corresponding to these feature dimensions (such as road density and terrain undulation) are identified as key constraining features. Information such as the names of these key constraining features, their contribution gradient values, and the actual values ​​of the corresponding original features are compiled to form a structured diagnostic result, providing precise targeting for subsequent planning scheme development.

[0171] In an optional embodiment, evacuation space planning is obtained through the following method:

[0172] S51. Analyze the diagnostic results. If the key limiting features involve the accessibility of emergency shelters, locate the geographical location of the relevant vulnerable units. Based on slope data, historical disaster data, and land use data, conduct a multi-criteria suitability analysis in the surrounding buffer zone to generate site selection recommendations.

[0173] Optionally, after analyzing the diagnostic results to identify key constraints related to the accessibility of emergency shelters, the geographic coordinates of the corresponding vulnerable units are obtained through spatial positioning. A reasonable surrounding buffer zone is delineated centered on the vulnerable unit, and the buffer zone range is determined in conjunction with the time threshold for pedestrian evacuation in mountainous areas. Based on slope data, areas with gentle terrain (slope below a preset safety threshold) are selected. Based on historical disaster data, areas with a high incidence of historical disasters are avoided (maintaining a preset safety distance from historical disaster points). Land use data is referenced to prioritize compatible land uses such as idle land and public green spaces. Through multi-criteria suitability analysis, the plots in the buffer zone are scored for suitability, and plots with scores higher than the suitability threshold are selected as candidate sites for shelters. Site selection suggestions are generated, including candidate site coordinates, suitability scores, and construction conditions.

[0174] S52. Analyze the diagnostic results. When the key constraint features involve the connectivity of the road network, based on the topological connection relationship of the unit graph structure, use network analysis methods to identify the key missing road segments or bottleneck road segments that restrict the connectivity of weak units, and generate road planning suggestions.

[0175] Optionally, based on key constraints related to road network connectivity, the topological connections between nodes and adjacent nodes corresponding to the target weak unit in the unit graph structure are extracted to construct a road network topology map. Using accessibility analysis and bottleneck identification algorithms from network analysis methods, the accessibility index from the weak unit node to surrounding refuge areas and main road nodes is calculated, and connectivity breaks with accessibility indices below a preset standard are located. Simultaneously, bottleneck road sections (such as narrow bridges and steep slopes) with traffic capacity below the emergency demand threshold are identified. Combining terrain conditions and existing road distribution, the routes of key missing road sections to be added and the scope of bottleneck road sections to be modified are determined, generating road planning suggestions including road section routes, construction standards, and modification priorities.

[0176] S53. Integrate site selection recommendations with road planning recommendations to form an evacuation space plan.

[0177] Optionally, site selection recommendations and road planning recommendations can be integrated to ensure that newly added shelters and planned roads form an efficient evacuation network. The spatial compatibility between candidate shelters and planned roads should be verified, road routes adjusted to ensure accessibility of shelters, and the overlap between the service area of ​​shelters and road coverage optimized. After integration, the spatial coordinates, technical parameters, implementation priorities, construction periods, and responsible entities for each planning project should be clearly defined, forming a systematic evacuation space plan to ensure the feasibility and emergency practicality of the planning scheme.

[0178] In an optional embodiment, dynamic path decision recommendations are obtained through the following method:

[0179] S61. After receiving a specific natural disaster early warning information, adjust the weights or states of relevant features in the multi-source heterogeneous data according to the disaster type, and re-execute S2 to S3 to obtain the updated dynamic evaluation score; the updated dynamic evaluation score is used to characterize the dynamic emergency evacuation capacity evaluation score under the disaster scenario.

[0180] Optionally, upon receiving early warning information for specific natural disasters such as earthquakes, landslides, and debris flows, the weights or states of core correlated features in the multi-source heterogeneous data are adjusted according to the impact characteristics of different disaster types. For example, for landslide disasters, the weights of topographic features such as terrain relief and slope aspect are increased; for debris flow disasters, the impact weights of features such as river distribution and historical disaster sites are strengthened; feature data that have been affected by disasters or have potential risks are marked as risk states. Based on the adjusted multi-source heterogeneous data, the unit graph structure construction and dual-flow feature enhancement graph convolutional network model evaluation process is re-executed, and a dynamic emergency evacuation capability evaluation score adapted to the current disaster scenario is output.

[0181] S62. Based on the updated dynamic assessment score, identify the current risk area and obtain the population distribution information and real-time capacity status of emergency shelters within the corresponding area.

[0182] Optionally, a risk threshold is set under disaster scenarios, and spatial units with dynamic assessment scores below this threshold are designated as current risk areas. The spatial extent and boundaries of the risk areas are extracted using GIS tools. Real-time population distribution information within the risk areas is retrieved simultaneously, including the number of permanent residents, the number of temporary migrants, and the age structure of the population. The real-time capacity status of each emergency shelter is obtained through the emergency management platform, including core information such as remaining capacity, number of people already accommodated, and remaining material reserves, providing basic data support for route planning.

[0183] S63. Taking residential areas within the risk zone as the starting point of the route and available emergency shelters as the ending point, the dynamic evaluation score is used as a constraint on the route's traffic resistance. The shortest path search algorithm is then used to calculate and generate recommended evacuation routes.

[0184] Optionally, the starting point of the route planning can be clearly defined as each residential area within the risk zone, and the endpoint as an available emergency shelter whose real-time capacity meets the reception needs. Dynamic evaluation scores are mapped to path traffic resistance coefficients; lower scores indicate greater traffic resistance. A comprehensive path cost function is constructed, incorporating traffic resistance, path distance, and road capacity. An improved A* shortest path search algorithm is employed, aiming to minimize the comprehensive path cost, to search for and generate optimal recommended evacuation routes from each residential area to the corresponding available shelter, while avoiding dangerous road sections with excessively high traffic resistance.

[0185] S64. Package the recommended evacuation routes with the estimated evacuation time and route capacity assessment information to generate dynamic route decision suggestions.

[0186] Optionally, based on the length of the recommended evacuation routes, road capacity, and population density within the risk area, the estimated evacuation time for each route is calculated; the route capacity is assessed using parameters such as road width and traffic efficiency to determine if there is a risk of congestion. The spatial orientation, nodes along the recommended evacuation routes, estimated evacuation time, route capacity, evacuation precautions, and information on corresponding refuge locations are integrated and packaged to form a structured dynamic route decision-making recommendation.

[0187] The aforementioned integrated assessment and planning method for emergency evacuation decision-making in mountainous villages and towns achieves accurate assessment of emergency evacuation capabilities through standardized processing of multi-source heterogeneous data and construction of superpixel unit graphs, combined with dual-stream feature-enhanced graph convolutional networks. It targets and locates key constraint features based on gradient contribution analysis, and generates targeted evacuation space planning that coordinates refuge sites and road networks. Furthermore, it can dynamically adjust parameters based on disaster early warnings to output evacuation route suggestions adapted to real-time risks. This effectively solves problems such as the crudeness of traditional emergency evacuation assessments in mountainous villages and towns, the ambiguity in the diagnosis of constraint factors, the insufficient targeting of planning schemes, and the lag in route decision-making under disaster scenarios. It improves the scientific nature and timeliness of emergency evacuation system construction, providing reliable technical support for disaster prevention, mitigation, and emergency management in mountainous villages and towns.

[0188] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0189] Based on the same inventive concept, this application also provides an integrated assessment and planning device for emergency evacuation decision-making in mountainous villages and towns, used to implement the integrated assessment and planning method for emergency evacuation decision-making in mountainous villages and towns described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the integrated assessment and planning device for emergency evacuation decision-making in mountainous villages and towns provided below can be found in the limitations of the integrated assessment and planning method for emergency evacuation decision-making in mountainous villages and towns described above, and will not be repeated here.

[0190] In one exemplary embodiment, such as Figure 4 The diagram shows a structural schematic of an integrated assessment and planning emergency evacuation decision-making device 20 for mountainous villages and towns, including:

[0191] The data acquisition and standard preprocessing module 21 is used to acquire multi-source heterogeneous data of the target mountainous villages and towns, and to perform standard preprocessing on the multi-source heterogeneous data to obtain multi-source standardized raster data; the multi-source heterogeneous data includes multimodal remote sensing images, digital elevation model data, geographic information vector data, disaster historical data and planning attribute data;

[0192] The cell graph construction module 22 is used to construct a cell graph structure with spatial cells as nodes and spatial adjacency relationships as edges based on multi-source standardized raster data.

[0193] The evacuation capability assessment module 23 is used to input the unit graph structure into the preset dual-stream feature-enhanced graph convolutional network model and output the emergency evacuation capability assessment score of each spatial unit in the unit graph structure; the dual-stream feature-enhanced graph convolutional network model includes a visual terrain feature branch, a semantic attribute feature branch, a cross-stream attention interaction module, and a multilayer perceptron classifier;

[0194] The constraint diagnosis module 24 is used to diagnose key constraint characteristics based on the emergency evacuation capability assessment score and obtain diagnostic results.

[0195] The evacuation space planning scheme generation module 25 is used to generate evacuation space planning schemes based on diagnostic results; the evacuation space planning schemes include evacuation space planning and dynamic route decision suggestions.

[0196] Furthermore, the unit diagram construction module 22 is also used for:

[0197] S11. Perform superpixel segmentation on the optical image in the multimodal remote sensing image, divide it into multiple spatially continuous and feature-similar superpixel regions, define each superpixel region as a spatial unit node, and obtain a node set.

[0198] S12. Extract visual appearance features and terrain features from the multi-source standardized raster data within the geographic space corresponding to each spatial unit node in the node set.

[0199] S13. Overlay and analyze the spatial location and geographic information vector data, disaster history data and planning attribute data of each spatial unit node to extract geographic semantic features, including road density and building density.

[0200] S14. The visual appearance features, terrain features and geographic semantic features corresponding to each spatial unit node are spliced ​​and fused to generate the fused feature vector of each spatial unit node, and the fused feature vectors of all spatial unit nodes are used to form a node feature matrix.

[0201] S15. Based on the spatial adjacency relationship of all spatial unit nodes, establish undirected edges between adjacent nodes to form the edge set of the graph, and construct the unit graph structure based on the node set, edge set, and node feature matrix.

[0202] Furthermore, the evacuation capacity assessment module 23 is also used for:

[0203] S21. Input the node feature matrix in the unit graph structure into the dual-stream feature enhancement graph convolutional network model, and split the fused feature vector of each spatial unit node into visual terrain feature sub-vector and semantic attribute feature sub-vector.

[0204] S22. Input the visual terrain feature sub-vectors into the visual terrain feature branch to obtain the visual enhancement features; the visual terrain feature branch includes the first image convolutional layer;

[0205] S23. Input the semantic attribute feature sub-vectors into the semantic attribute feature branch to obtain the semantic enhancement features; the semantic attribute feature branch includes the second graph convolutional layer;

[0206] S24. Through the cross-stream attention interaction module, calculate the correlation weight between visual enhancement features and semantic enhancement features, and perform weighted fusion of visual enhancement features and semantic enhancement features based on the correlation weight to generate a unified deep feature representation for each spatial unit node.

[0207] S25. Input the unified deep feature representation into the multilayer perceptron classifier to obtain the emergency evacuation capability assessment score of each spatial unit node corresponding to the spatial unit.

[0208] Furthermore, the evacuation capacity assessment module 23 is also used for:

[0209] S31. Obtain multi-source heterogeneous data from multiple historical mountainous village and town cases to obtain a training graph sample set;

[0210] S32. For each training map sample in the training map sample set, use the comprehensive evaluation method combining the analytic hierarchy process and the entropy weight method to determine the weight of each indicator in the corresponding emergency evacuation capability evaluation index system, and combine the fuzzy comprehensive evaluation method to calculate the emergency evacuation capability score for each spatial unit in the training map sample, which serves as the supervised training label for the training map sample.

[0211] S33. Using the cell graph structure of the training image samples as input, the corresponding supervised training labels as the target output, and the mean squared error as the loss function, the parameters of the two-stream feature enhancement graph convolutional network model are iteratively optimized and trained using the backpropagation algorithm; where the expression for the mean squared error is:

[0212]

[0213] In the formula, Indicates mean square error. This represents the total number of spatial unit nodes in a training batch. Indicates the first Supervised training label scores corresponding to each spatial unit node The model represents the first The emergency evacuation capability assessment score predicted by each spatial unit node;

[0214] S34. When the model's performance on the validation set stabilizes, stop training, save the model parameters, and obtain the final two-stream feature-enhanced graph convolutional network model.

[0215] Furthermore, the constraint diagnostic module 24 is also used for:

[0216] S41. Based on the preset weak threshold, select units from all spatial units whose emergency evacuation capability assessment scores are lower than the weak threshold and mark them as target weak units.

[0217] S42. For each target weak unit, obtain the original fusion feature vector of the corresponding node in the unit graph structure;

[0218] S43. Calculate the gradient value of the contribution of each feature dimension in the fused feature vector to the emergency evacuation capability assessment score;

[0219] S44. Sort the data according to the absolute size of the contribution gradient values, identify the feature dimensions that have the main negative impact on the emergency evacuation capability assessment score, determine the original data features represented by the feature dimensions as key constraint features, and summarize to generate diagnostic results.

[0220] Furthermore, the evacuation space planning scheme generation module 25 is also used for:

[0221] S51. Analyze the diagnostic results. If the key limiting features involve the accessibility of emergency shelters, locate the geographical location of the relevant vulnerable units. Based on slope data, historical disaster data, and land use data, conduct a multi-criteria suitability analysis in the surrounding buffer zone to generate site selection recommendations.

[0222] S52. Analyze the diagnostic results. When the key constraint features involve the connectivity of the road network, based on the topological connection relationship of the unit graph structure, use network analysis methods to identify the key missing road segments or bottleneck road segments that restrict the connectivity of weak units and generate road planning suggestions.

[0223] S53. Integrate site selection recommendations with road planning recommendations to form an evacuation space plan.

[0224] Furthermore, the evacuation space planning scheme generation module 25 is also used for:

[0225] S61. After receiving a specific natural disaster early warning information, adjust the weights or states of relevant features in the multi-source heterogeneous data according to the disaster type, and re-execute S2 to S3 to obtain the updated dynamic evaluation score; the updated dynamic evaluation score is used to characterize the dynamic emergency evacuation capacity evaluation score under the disaster scenario.

[0226] S62. Based on the updated dynamic assessment score, identify the current risk area and obtain the population distribution information and real-time capacity status of emergency shelters within the corresponding area.

[0227] S63. Taking residential areas within the risk zone as the starting point of the route and available emergency shelters as the ending point, the dynamic evaluation score is used as a constraint on the route's traffic resistance. The shortest path search algorithm is then used to calculate and generate recommended evacuation routes.

[0228] S64. Package the recommended evacuation routes with the estimated evacuation time and route capacity assessment information to generate dynamic route decision suggestions.

[0229] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the integrated assessment and planning emergency evacuation decision-making method for mountainous villages and towns as described above.

[0230] In one embodiment, a computer-readable storage medium is provided, which stores various media capable of storing program code, such as a computer program, a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk. When the computer program is executed by a processor, it implements the steps in the above-described method embodiments.

[0231] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0232] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for emergency evacuation decision-making in mountainous villages and towns that integrates assessment and planning, characterized in that, The method includes: S1. Acquire multi-source heterogeneous data of target mountain villages and towns, and perform standardized preprocessing on the multi-source heterogeneous data to obtain multi-source standardized raster data; the multi-source heterogeneous data includes multimodal remote sensing images, digital elevation model data, geographic information vector data, disaster historical data and planning attribute data; S2. Based on the multi-source standardized raster data, construct a cell graph structure with spatial cells as nodes and spatial adjacency relationships as edges; S3. Input the unit graph structure into a preset dual-stream feature-enhanced graph convolutional network model, and output the emergency evacuation capability evaluation score of each spatial unit in the unit graph structure; the dual-stream feature-enhanced graph convolutional network model includes a visual terrain feature branch, a semantic attribute feature branch, a cross-stream attention interaction module, and a multilayer perceptron classifier; S4. Based on the emergency evacuation capability assessment score, diagnose key limiting characteristics and obtain diagnostic results; S5. Generate an evacuation space planning scheme based on the diagnostic results; the evacuation space planning scheme includes evacuation space planning and dynamic path decision suggestions.

2. The method according to claim 1, characterized in that, The construction of a cell graph structure based on the multi-source standardized raster data, with spatial cells as nodes and spatial adjacency relationships as edges, includes: S11. Perform superpixel segmentation on the optical image in the multimodal remote sensing image to divide it into multiple spatially continuous and feature-similar superpixel regions. Define each superpixel region as a spatial unit node to obtain a node set. S12. Extract visual appearance features and terrain features from the multi-source standardized raster data within the geographic space corresponding to each spatial unit node in the node set. S13. Overlay and analyze the spatial location of each spatial unit node with the geographic information vector data, the disaster history data, and the planning attribute data to extract geographic semantic features, including road density and building density. S14. The visual appearance features, terrain features and geographic semantic features corresponding to each spatial unit node are spliced ​​and fused to generate the fused feature vector of each spatial unit node, and the fused feature vectors of all spatial unit nodes are used to form a node feature matrix. S15. Based on the spatial adjacency relationship of all the spatial unit nodes, establish undirected edges between adjacent nodes to form the edge set of the graph, and construct the unit graph structure based on the node set, the edge set, and the node feature matrix.

3. The method according to claim 2, characterized in that, The step of inputting the unit graph structure into a preset dual-stream feature-enhanced graph convolutional network model and outputting an emergency evacuation capability evaluation score for each spatial unit in the unit graph structure includes: S21. Input the node feature matrix in the unit graph structure into the dual-stream feature enhancement graph convolutional network model, and split the fused feature vector of each spatial unit node into visual terrain feature sub-vector and semantic attribute feature sub-vector; S22. Input the visual terrain feature sub-vector into the visual terrain feature branch to obtain visual enhancement features; the visual terrain feature branch includes a first graph convolutional layer. S23. Input the semantic attribute feature sub-vector into the semantic attribute feature branch to obtain semantic enhancement features; the semantic attribute feature branch includes a second graph convolutional layer; S24. Through the cross-stream attention interaction module, calculate the correlation weight between the visual enhancement feature and the semantic enhancement feature, and perform weighted fusion of the visual enhancement feature and the semantic enhancement feature based on the correlation weight to generate a unified depth feature representation for each spatial unit node. S25. Input the unified depth feature representation into the multilayer perceptron classifier to obtain the emergency evacuation capability evaluation score of the spatial unit corresponding to each spatial unit node.

4. The method according to claim 1, characterized in that, The training of the dual-stream feature enhancement map convolutional network model is obtained through the following method: S31. Obtain the multi-source heterogeneous data of multiple historical mountainous village and town cases to obtain a training graph sample set; S32. For each training map sample in the training map sample set, the weight of each indicator in the corresponding emergency evacuation capability evaluation index system is determined by a comprehensive evaluation method combining the analytic hierarchy process and the entropy weight method. Then, the emergency evacuation capability score is calculated for each spatial unit in the training map sample by combining the fuzzy comprehensive evaluation method, which serves as the supervised training label for the training map sample. S33. Using the cell graph structure of the training image samples as input, the corresponding supervised training labels as the target output, and the mean squared error as the loss function, the parameters of the dual-stream feature enhancement graph convolutional network model are iteratively optimized and trained using the backpropagation algorithm; wherein, the expression for the mean squared error is: In the formula, Indicates mean square error. This represents the total number of spatial unit nodes in a training batch. Indicates the first Supervised training label scores corresponding to each spatial unit node The model represents the first The emergency evacuation capability assessment score predicted by each spatial unit node; S34. When the model's performance on the validation set tends to stabilize, stop training, save the model parameters, and obtain the final dual-stream feature enhancement graph convolutional network model.

5. The method according to claim 1, characterized in that, The diagnostic results, obtained by diagnosing key limiting characteristics based on the emergency evacuation capability assessment score, include: S41. Based on the preset weak threshold, select the units whose emergency evacuation capability assessment scores are lower than the weak threshold from all spatial units and mark them as target weak units. S42. For each target weak unit in the unit graph structure, obtain the original fusion feature vector of the corresponding node; S43. Calculate the gradient value of the contribution of each feature dimension in the fused feature vector to the emergency evacuation capability assessment score; S44. Sort the contribution gradient values ​​according to their absolute magnitude, identify the feature dimensions that have a major negative impact on the emergency evacuation capability assessment score, determine the original data features represented by the feature dimensions as the key constraint features, and summarize to generate the diagnostic results.

6. The method according to claim 1, characterized in that, The evacuation space planning was obtained through the following method: S51. Analyze the diagnostic results. If the key constraint features involve the accessibility of emergency shelters, locate the geographical location of the relevant vulnerable units. Based on slope data, historical disaster data, and land use data, conduct a multi-criteria suitability analysis in the surrounding buffer zone to generate site selection recommendations. S52. Analyze the diagnostic results. When the key constraint features involve the connectivity of the road network, based on the topological connection relationship of the unit graph structure, use network analysis methods to identify the key missing road segments or bottleneck road segments that restrict the connectivity of the weak unit and generate road planning suggestions. S53. Integrate the site selection recommendations with the road planning recommendations to form the evacuation space plan.

7. The method according to claim 6, characterized in that, The dynamic path decision suggestion is obtained through the following method: S61. After receiving a specific natural disaster early warning information, adjust the weights or states of relevant features in the multi-source heterogeneous data according to the disaster type, and re-execute S2 to S3 to obtain an updated dynamic evaluation score; the updated dynamic evaluation score is used to characterize the dynamic emergency evacuation capacity evaluation score under the disaster scenario. S62. Based on the updated dynamic assessment score, identify the current risk area and obtain the population distribution information and real-time capacity status of emergency shelters within the corresponding area. S63. Taking the residential areas within the risk area as the starting point of the path and the available emergency shelters as the ending point, the dynamic evaluation score is used as a constraint on the path's traffic resistance. The shortest path search algorithm is then used to calculate and generate a recommended evacuation path. S64. Package the recommended evacuation route with the estimated evacuation time and route capacity assessment information to generate the dynamic route decision suggestion.

8. An integrated assessment and planning emergency evacuation decision-making device for mountainous villages and towns, characterized in that, The device includes: The data acquisition and standard preprocessing module is used to acquire multi-source heterogeneous data of target mountain villages and towns, and to perform standardization preprocessing on the multi-source heterogeneous data to obtain multi-source standardized raster data; the multi-source heterogeneous data includes multimodal remote sensing images, digital elevation model data, geographic information vector data, disaster historical data, and planning attribute data; The cell graph construction module is used to construct a cell graph structure with spatial cells as nodes and spatial adjacency relationships as edges based on the multi-source standardized raster data. The evacuation capability assessment module is used to input the unit graph structure into a preset dual-stream feature-enhanced graph convolutional network model and output the emergency evacuation capability assessment score of each spatial unit in the unit graph structure; the dual-stream feature-enhanced graph convolutional network model includes a visual terrain feature branch, a semantic attribute feature branch, a cross-stream attention interaction module, and a multilayer perceptron classifier; The constraint diagnosis module is used to diagnose key constraint features based on the emergency evacuation capability assessment score and obtain diagnostic results. An evacuation space planning scheme generation module is used to generate an evacuation space planning scheme based on the diagnostic results; the evacuation space planning scheme includes evacuation space planning and dynamic path decision suggestions.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images