A Graph Neural Network-Based Method for Assessing Storm Surge Resilience in Coastal Cities

Abstract

This invention discloses a method for assessing the resilience of coastal cities to storm surge disasters based on graph neural networks. The method involves collecting multi-source heterogeneous data related to storm surge disasters in coastal cities, including geographic information data, infrastructure data, and historical storm surge data. Based on the geographic information data and infrastructure data, a coastal city infrastructure network structure data is constructed, and then a graph neural network model is built. The feature parameters of the historical storm surge data are fused with the coastal city infrastructure network structure data and input into the trained graph neural network model to simulate the hierarchical impact of storm surge disasters on the nodes of the coastal city infrastructure. Based on the graph neural network propagation results, the resilience of each node in the coastal city infrastructure network under the influence of storm surges is assessed. Based on the obtained resilience assessment index data of each node, different functional areas are summarized and analyzed to form an overall coastal city resilience assessment result.

Description

Technical Field

[0001] This invention belongs to the technical field of meteorological disaster risk prediction and assessment for coastlines, specifically involving a method for assessing the resilience of coastal cities to storm surge disasters based on graph neural networks. Background Technology

[0002] With the intensification of global climate change, frequent storm surges have become one of the natural disasters that seriously threaten the public safety and operational stability of infrastructure systems in coastal cities. Storm surges often lead to multiple cascading failures, such as damage to coastal seawalls, disruption of port operations, and exacerbation of urban flooding, resulting in decreased urban resilience and difficulties in post-disaster recovery. Therefore, how to conduct resilience modeling, impact propagation analysis, and assessment optimization under storm surge disasters based on complex urban infrastructure networks has become a research hotspot in the field of disaster prevention and mitigation in coastal cities.

[0003] Traditional urban resilience assessment methods typically rely on expert experience or static indicator evaluation systems, which struggle to reflect the interactions between infrastructures and the dynamic propagation process of disaster impacts, lacking model universality and predictive capability. Therefore, there is an urgent need to design an efficient and accurate resilience assessment scheme for coastal cities under storm surge scenarios. Summary of the Invention

[0004] The present invention aims to overcome the shortcomings of the prior art by providing a method for assessing the resilience of coastal cities to storm surge disasters based on graph neural networks.

[0005] This invention is achieved through the following technical solution:

[0006] A method for assessing the resilience of coastal cities to storm surge disasters based on graph neural networks includes the following steps:

[0007] Step 1: Collect multi-source heterogeneous data related to storm surge disasters in coastal cities, including: geographic information data, infrastructure data, and historical storm surge data;

[0008] Step 2: Preprocess the data collected in Step 1;

[0009] Step 3: Construct infrastructure network structure data for coastal cities based on geographic information data and infrastructure data;

[0010] Step 4: Construct a graph neural network model based on the constructed coastal city infrastructure network structure data;

[0011] Step 5: After fusing the feature parameters of historical storm surge data with the coastal city infrastructure network structure data, input the data into the trained graph neural network model; based on the cascading impact propagation mechanism of the graph neural network model, simulate the step-by-step impact of storm surge disasters on the nodes of coastal city infrastructure, and obtain the dynamic changes of the temporal state of each node in the coastal city infrastructure network under the impact of storm surge.

[0012] Step 6: Based on the graph neural network propagation results in Step 5, conduct a resilience assessment of the state evolution of each node in the coastal city infrastructure network under the influence of storm surge; based on the obtained resilience assessment index data of each node, summarize and analyze different functional areas to form an overall coastal city resilience assessment result.

[0013] In the above technical solution, in step 1, the geographic information data includes high-resolution urban topography, geomorphological data and coastline shape; infrastructure data includes port facility layout, seawall structural parameters and residential area distribution data; historical storm surge data includes the occurrence time, tide height, path trajectory, duration, wind speed and direction at landfall and central pressure value of historical storm surges covering many years.

[0014] In the above technical solution, step 2, the preprocessing includes: outlier removal, missing value filling, format unification, and standardization.

[0015] In the above technical solution, in step 3, entities such as ports, seawalls, and residential areas are abstracted as nodes, and edge relationships are established based on their spatial proximity or functional association to construct a coastal urban infrastructure network map, wherein: residential areas are connected by roads; edges are established between ports and the infrastructure that supports their transportation; and edges are established between seawalls and the areas protected by them.

[0016] In the above technical solution, step 4 includes:

[0017] Step 4.1: First, define the node and edge features;

[0018] Step 4.2: A graph neural network model based on a supervised graph sparsification mechanism is adopted. For the infrastructure network structure of coastal cities, label information is introduced to guide edge sampling, constructing a task-related sparse subgraph. The model design includes the following steps: ① Edge probability learning: By introducing structural priors and label supervision information, the importance of each edge is learned; ② Sparse subgraph sampling: A sparse subgraph is generated based on the edge probability distribution, retaining the key structures that have the greatest impact on the evaluation results; ③ Graph neural network training: Node representation learning and resilience prediction are performed on the sparse graph structure.

[0019] In the above technical solution, in step 4.1, a distinctive feature vector is defined for each type of node, including: residential area nodes: building density, total population, average building age, and shelter coverage; port nodes: annual throughput, cargo type, and facility disaster resistance level; seawall nodes: seawall length, height, structural type, service life, and maintenance cycle; edge features include: connection strength, connection reliability, and response time.

[0020] In the above technical solution, in step 4.2, for edge probability learning: combined with node feature embedding, a multilayer perceptron (MLP) is used to extract the correlation score of the edge.

[0021] In the above technical solution, in step 4.2, for sparse subgraph sampling: to ensure that the sampling process is supported by gradient optimization, Gumbel noise is introduced to achieve approximately discrete sampling; to construct a differentiable sparse edge sampling distribution, Softmax normalization is introduced, combined with a temperature annealing mechanism and temperature parameters. By controlling the smoothness of the distribution, a dynamic balance between exploration and determinism can be achieved.

[0022] In the above technical solution, in step 4.2, for graph neural network training: the METIS graph partitioning algorithm is used to divide the original large graph into several local subgraphs, realizing distributed batch training and effectively reducing memory overhead; the overall training process is divided into three stages: teacher model training, student model distillation learning, and inference deployment.

[0023] In the above technical solution, step 6 includes the following resilience assessment indicators for each node:

[0024] Node recovery time: refers to the time required for a node's functionality to recover from a damaged state to a normal state;

[0025] Node function loss rate: refers to the proportion of node function that decreases during the impact of a disaster;

[0026] Node connectivity interruption time: refers to the duration of interruption of communication or traffic between nodes due to cascading impacts.

[0027] The advantages and beneficial effects of this invention are as follows:

[0028] This invention can effectively simulate the cascading impact process of complex infrastructure networks in coastal cities under disaster shocks, quantify system resilience indicators, and generate optimized disaster prevention strategies.

[0029] This invention proposes a graph neural network model based on a supervised graph sparsification mechanism: it uses label information to guide the edge sampling process, dynamically generating sparse subgraphs related to the prediction task from the whole graph, which reduces computational overhead while retaining key structural information and adapts to the needs of large-scale heterogeneous graph modeling.

[0030] This invention constructs a probability distribution of edge existence, and by fusing node embedding features and introducing degree priors, accurately estimates the retention importance of each edge in the evaluation task.

[0031] This invention proposes a sparse subgraph sampling method that combines Gumbel noise and temperature annealing: while preserving the differentiability of the model, it achieves a controllable exploration-utilization balance, so that the sampling results are both diverse and approach the optimal structure.

[0032] This invention proposes a label attention distillation mechanism: constructing a teacher-student framework, introducing graph-level label embedding, and guiding node representations to align with graph-level labels through an attention mechanism, which significantly improves the semantic consistency of node representations in classification tasks. Attached Figure Description

[0033] Figure 1 This is a flowchart of the storm surge disaster resilience assessment method for coastal cities based on graph neural networks, as proposed in this invention.

[0034] For those skilled in the art, other related figures can be obtained from the above figures without any creative effort. Detailed Implementation

[0035] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described below with reference to specific embodiments.

[0036] This invention proposes a method for assessing the storm surge resilience of coastal cities based on graph neural networks. (See appendix.) Figure 1 The specific steps are as follows.

[0037] Step 1: Collect multi-source heterogeneous data related to storm surge disasters in coastal cities. Specifically, this includes:

[0038] Geographic information data: including high-resolution urban topography, geomorphological data, and coastline shapes.

[0039] Infrastructure data includes port facility layout (such as wharf location, number of berths and throughput capacity), seawall structural parameters (such as height, width, material, and design wave resistance level), and residential area distribution (such as building density, population density, building type and year of construction).

[0040] Historical storm surge data includes the occurrence time, tide height, path, duration, wind speed and direction at landfall, and central pressure value of historical storm surges covering many years.

[0041] The data sources include government geographic information databases, real-time data platforms of marine monitoring stations, meteorological records, and urban planning archives.

[0042] Step 2: Preprocess the data collected in Step 1.

[0043] Specifically, the preprocessing includes:

[0044] Outlier removal: Identify and remove noisy data caused by equipment failure or transmission errors.

[0045] Missing value imputation: Linear interpolation, spline interpolation, or K-nearest neighbor algorithms are used for imputation.

[0046] Standardized format: Convert data from different sources and coordinate systems into a unified format and perform coordinate calibration.

[0047] Standardization processing: Normalize all types of data to the [0,1] interval for subsequent model training.

[0048] Step 3: Construct infrastructure network structure data for coastal cities based on geographic information data and infrastructure data.

[0049] Specifically, based on geographic information data and infrastructure data, entities such as ports, seawalls, and residential areas are abstracted as nodes in the graph. Edge relationships are established based on their spatial proximity or functional association to construct a coastal city infrastructure network graph. Among them: residential areas are connected by roads; edges are established between ports and the infrastructure (roads, railways) that support their transportation; and edges are established between seawalls and the areas protected by them.

[0050] Edge properties include connectivity strength (e.g., traffic flow), reliability (e.g., wind and wave resistance), and response time (e.g., the shortest travel time from a residential area to a shelter). The resulting graph is stored in the form of an adjacency matrix, an edge list, etc., to meet the requirements of graph neural network modeling.

[0051] Step 4: Based on the constructed coastal city infrastructure network structure data, construct a graph neural network (GNN) model.

[0052] Step 4.1: Define node and edge features.

[0053] Define a discriminative feature vector for each type of node, including but not limited to:

[0054] 1. Residential area nodes: building density, total population, average age of buildings, and coverage of refuge areas;

[0055] 2. Port nodes: annual throughput, cargo types, and facility disaster resistance level;

[0056] 3. Seawall details: seawall length, height, structural type, service life, maintenance cycle, etc.

[0057] Edge characteristics include: connection strength, connection reliability, response time, etc.

[0058] The aforementioned features constitute the input to the graph neural network model, used to characterize the interaction effects between nodes and structural fragility.

[0059] Step 4.2: Model selection and design.

[0060] A graph neural network model based on supervised graph sparsification mechanism is adopted. For the highly heterogeneous infrastructure network structure of coastal cities, label information is introduced to guide edge sampling, thereby constructing a task-related sparse subgraph, which can effectively reduce the computational complexity of the model and improve the prediction performance.

[0061] Model design includes the following core processes:

[0062] 1. Edge probability learning: By introducing structural priors and label supervision information, the importance of each edge is learned;

[0063] 2. Sparse subgraph sampling: Generate sparse subgraphs based on edge probability distribution, retaining the key structures that have the greatest impact on the evaluation results;

[0064] 3. Graph Neural Network Training: Node representation learning and resilience prediction on sparse graph structures.

[0065] This closed-loop structure realizes an adaptive graph modeling mechanism of "structure learning - efficient modeling - performance optimization", which is suitable for handling large-scale, low-homogeneity complex urban network structures.

[0066] Specifically, step 4.2 includes the following steps:

[0067] Step 4.2.1: Perform edge probability learning.

[0068] In the task of constructing storm surge resilience assessment, the edges used for graph sparsification have a probability distribution. By using the characteristic relationships between nodes, the importance of each edge being retained in the construction of the sparse subgraph is estimated.

[0069] By combining node feature embedding, a multilayer perceptron (MLP) is used to extract the correlation score of edges. For any edge... Its corresponding node and The embedding vector is and The edge weights are calculated by fusing features through concatenation and element-wise multiplication. :

[0070] ;

[0071] , : are nodes respectively and The embedding vector;

[0072] Vector concatenation operation;

[0073] Element-wise multiplication of vectors;

[0074] Activation functions (such as Sigmoid) ensure that edge weights fall within the [0,1] interval;

[0075] Multilayer perceptron is used to model nonlinear feature interaction relationships.

[0076] To avoid high-degree nodes occupying all sampling edges, a prior distribution based on node degree is introduced to guide the model to focus more on the connections between low-degree nodes, thus avoiding the loss of key connections.

[0077] ;

[0078] :node and The degree (i.e., the number of connected edges);

[0079] :side The prior sampling tendency encourages the preservation of potentially critical but sparse connections in edge structures.

[0080] A relationship that is directly proportional to.

[0081] Step 4.2.2: Perform sparse subgraph sampling.

[0082] This step is based on the probability distribution of the edges. Sampling sparse subgraphs that preserve key structures from the original infrastructure graph This reduces the computational complexity of graph neural networks and improves the model's generalization performance in storm surge scenarios. From the edge probability distribution, according to the sparsity... (e.g., 20%) Sample a fixed number of edges to form a subgraph structure:

[0083] ;

[0084] : Sparsity retention ratio;

[0085] : The total number of edges in the original graph;

[0086] : Actual number of sampled edges.

[0087] To enable gradient optimization for the sampling process, Gumbel noise is introduced to achieve approximately discrete sampling. To construct a differentiable sparse edge sampling distribution, Softmax normalization is introduced, and temperature annealing and temperature parameters are used to control the smoothness of the distribution, thereby achieving a dynamic balance between exploratory and deterministic approaches.

[0088] ;

[0089] :side The normalized probability;

[0090] :side The Gumbel noise improves the randomness of sampling;

[0091] The logarithm of the edge weights helps to enhance numerical stability;

[0092] The set of all edges in the graph;

[0093] Temperature parameter controls the smoothness of the distribution. Approaching hard sampling, (Approaching a uniform distribution).

[0094] To ensure the model training is highly exploratory in the early stages and more deterministic in the later stages, a temperature annealing mechanism is introduced, with the following formula:

[0095] ;

[0096] ;

[0097] Initial temperature;

[0098] Minimum temperature;

[0099] Annealing rate, controlling the rate of temperature decrease;

[0100] : Current training round number;

[0101] max_epochs: Total number of epochs;

[0102] : Maximum value operation.

[0103] Through theoretical analysis and empirical verification, it is ensured that the sparse subgraph obtained by sampling retains the key edges of the original graph structure without destroying the connectivity of key paths and nodes, and has a high structural similarity with the "ideal subgraph".

[0104] Step 4.2.3: Training the graph neural network.

[0105] This step aims to perform node classification based on the sampled sparse subgraphs and optimize model parameters accordingly to achieve efficient learning and representation of graph structure information. Considering the scalability issues of large graph processing, the METIS graph partitioning algorithm is used to divide the original large graph into several local subgraphs, enabling distributed batch training and effectively reducing memory overhead.

[0106] To improve the model's discriminative performance in graph classification tasks, this step introduces a label attention distillation mechanism, integrating label guidance and distillation learning strategies to effectively alleviate the problem of insufficient node embedding and graph-level label alignment in traditional GNNs. The overall training process is divided into three stages: teacher model training, student model distillation learning, and inference deployment, as detailed below:

[0107] Step 4.2.3.1: Teacher model training.

[0108] The GNN backbone network adopts mainstream GNN architectures such as GCN and GAT. The GNN layer update formula is:

[0109] ;

[0110] : No. Layer nodes The embedding vector;

[0111] :node The set of neighbors;

[0112] Neighbor aggregation functions (such as mean, attention mechanism);

[0113] Node update functions (such as MLP + activation functions);

[0114] : The set of all nodes in the graph.

[0115] Through the adjacency matrix and node features Constructing the initial node embedding:

[0116] ;

[0117] The adjacency matrix of a graph;

[0118] Node feature matrix (initial features);

[0119] Learnable parameters of the GNN backbone;

[0120] The overall mapping function of the GNN backbone outputs the node embedding matrix. .

[0121] The label attention distillation mechanism is introduced to construct an ideal embedding. The label attention encoder utilizes an MLP network to process graph labels. Perform embedding to obtain the label embedding vector: ;

[0122] :picture The label embedding vector;

[0123] :picture Unique hot tags;

[0124] Label encoders (such as MLPs) map labels to the latent space;

[0125] Graph dataset.

[0126] By leveraging a Transformer-style attention mechanism, node embeddings and label embeddings are combined to obtain an ideal node representation. :

[0127] ;

[0128] ;

[0129] ;

[0130] The tag embedding projection is the query vector;

[0131] Node embeddings are projected as key and value vectors;

[0132] Learnable projection matrix;

[0133] Attention temperature coefficient, controlling the sharpness of the Softmax distribution (adjusting exploration - utilizing balance);

[0134] Layer normalization helps mitigate internal covariate bias.

[0135] Feedforward neural networks introduce nonlinear transformations;

[0136] Ideal node embeddings generated by the teacher model (results after fusing label information).

[0137] Step 4.2.3.2: Student model distillation learning.

[0138] The student model uses the same GNN backbone structure and classification head as the teacher, inheriting label-guided node knowledge to ensure learning within the same semantic space.

[0139] During training, the teacher model parameters remain fixed, while the student model is minimized through backpropagation. We learn high-quality node embeddings with label guidance from the teacher model. Loss function. Multi-loss joint optimization is adopted:

[0140] ;

[0141] Weighting parameters are used to weigh the proportion of each loss.

[0142] Homogeneity loss encourages a higher probability of edge connections between nodes with the same label, thereby strengthening intra-class structural consistency. The formula is as follows:

[0143] ;

[0144] The indicator function is 1 if the condition is true, and 0 otherwise.

[0145] The set of all edges in the graph;

[0146] :side Edge weights;

[0147] For the set of label nodes, only the nodes with labels are computed. For edges between elements, penalize the low probability of edges with the same label;

[0148] :node and The true category label;

[0149] :when A value of 1 indicates that only nodes of the same type are considered when the labels are the same.

[0150] Consistency loss forces the weights of structural edges to remain consistent with the semantic similarity (such as cosine similarity) between node embeddings, thereby enhancing the coupling between the graph's structure and features.

[0151] ;

[0152] : No. Layer nodes Embedded cosine similarity;

[0153] : are nodes respectively and The embedding vector;

[0154] Cross-entropy loss function for graph-level label classification, measuring classification error:

[0155] ;

[0156] Number of training samples;

[0157] : No. The true label of each image (0 or 1);

[0158] Teacher model for the first The predicted probability of each graph;

[0159] Distillation loss, using the student model, measures the difference between teacher and student embeddings by minimizing the mean squared error loss between the student model and the ideal teacher embedding.

[0160] ;

[0161] The first teacher model generated The node embedding matrix of the graph;

[0162] The first student model generated The node embedding matrix of the graph;

[0163] : Squared Euclidean distance (i.e., mean square error).

[0164] Furthermore, during the inference phase, the trained student model is used to perform the graph classification task. Since the student model inherits the label-awareness ability from the teacher model through the distillation process, accurate classification can be achieved while maintaining low computational cost. The prediction process includes:

[0165] 1. Input subgraph , It is an adjacency matrix. The node feature matrix;

[0166] 2. Embedding of GNN backbone generation nodes , Node embedding matrix generated for the student model;

[0167] 3. Generate a graph-level representation: ;

[0168] : Graph-level representation of vectors; The set of all nodes in the graph; Pooling functions (such as mean pooling, max pooling, or attention pooling); :node The embedding vector;

[0169] 4. Class probabilities are output by the classification head:

[0170] ;

[0171] The graph represents the predicted probability distribution for each category. The weight matrix of the classification head; : The bias term of the category header;

[0172] 5. Finally, select the one with the highest probability as the prediction result:

[0173] ;

[0174] Total number of categories; : indicates the predicted probability that the graph belongs to class c; The final map category predicted by the model.

[0175] Step 5: After fusing the feature parameters of historical storm surge data with the coastal city infrastructure network structure data, input the data into the trained graph neural network model; based on the cascading impact propagation mechanism of the graph neural network model, simulate the step-by-step impact of storm surge disasters on the nodes of coastal city infrastructure, and obtain the dynamic changes of the temporal state of each node in the coastal city infrastructure network under the impact of storm surge.

[0176] Step 5.1: Input disaster scenario simulation data.

[0177] In this step, the characteristic parameters of historical storm surge data (storm surge occurrence time, tide height, path trajectory, duration, wind speed and direction at landfall, and central pressure) are fused with the coastal city infrastructure network structure data established in step 3, and then input into the graph neural network model trained in step 4 to construct storm surge impact scenarios. By inputting various storm surge intensity levels and path combinations, the system response behavior of coastal cities under different storm surge scenarios is simulated.

[0178] Step 5.2: Calculation of the cascading impact propagation mechanism based on the graph neural network model (GNN).

[0179] A specific information propagation mechanism constructed in a graph neural network model is employed to simulate the cascading impact of storm surge disasters on various types of urban infrastructure nodes. In the initial state of the model, the seawall node, as a coastal defense barrier, is the first to receive external disturbance input from the storm surge, and its damage degree is jointly determined by node characteristics (such as structure type, wave resistance level, etc.) and storm surge parameters. The model then transmits the damage information of the seawall node to downstream functional nodes such as port nodes and residential area nodes through an edge propagation mechanism. Each type of node updates its state characteristics according to its own attributes (such as the storage and transportation capacity of the port, the disaster resistance level of the residential area buildings, etc.) and continues to propagate the impact to other connected nodes, forming a cascading effect chain. Through multiple rounds of graph information propagation iterations, the dynamic changes of the temporal state of each node in the coastal urban infrastructure network under the impact of storm surge are obtained, which is used to characterize the cascading damage process of the system.

[0180] Step 6: Based on the graph neural network propagation results in Step 5, assess the resilience of the state evolution of each node in the coastal city infrastructure network under the influence of storm surge.

[0181] The resilience assessment metrics for each node include, but are not limited to, the following categories:

[0182] Node recovery time: refers to the time required for a node's functionality to recover from a damaged state to a normal state;

[0183] Node function loss rate: refers to the proportion of node function that decreases during the impact of a disaster;

[0184] Node connectivity interruption time: refers to the duration of interruption of communication or traffic between nodes due to cascading impacts.

[0185] Based on the resilience assessment indicators of each node, the results of the overall urban resilience assessment are summarized and analyzed for different functional areas (such as residential areas, ports, transportation hubs, etc.).

[0186] Furthermore, based on the assessment results of step 6, optimization strategies for urban resilience can be designed.

[0187] As a preferred approach, based on the evaluation results of step 6, a multi-objective intelligent optimization algorithm is used to construct a search model for urban storm surge resilience enhancement strategies. Optimization methods with global search capabilities, such as genetic algorithms and particle swarm optimization, are selected to generate the optimal combination of resilience enhancement strategies, considering resource constraints (such as fiscal budgets and spatial limitations). Taking a residential area as an example, strategy variables such as reinforcing building structures, increasing emergency refuge capacity, and optimizing drainage systems are encoded, and objective functions (such as minimizing average recovery time and reducing functional loss rate) are set. Through iterative search, the optimal strategy output under the constraints is obtained.

[0188] The optimization process can incorporate realistic feasibility constraints, including the timeline for strategy implementation and land use compatibility requirements, to ensure the generated results are executable in real-world scenarios. The final output is a set of engineering-feasible resilience enhancement schemes, providing technical support for the proactive defense and rapid recovery of coastal cities under high-risk storm surge scenarios.

[0189] The present invention has been described above by way of example. It should be noted that any simple modifications, alterations or other equivalent substitutions that can be made by those skilled in the art without creative effort without departing from the core of the present invention fall within the protection scope of the present invention.

Claims

1. A method for assessing the resilience of coastal cities to storm surge disasters based on graph neural networks, characterized in that: Includes the following steps: Step 1: Collect multi-source heterogeneous data related to storm surge disasters in coastal cities, including: geographic information data, infrastructure data, and historical storm surge data; Step 2: Preprocess the data collected in Step 1; Step 3: Construct infrastructure network structure data for coastal cities based on geographic information data and infrastructure data; Step 4: Based on the constructed coastal city infrastructure network structure data, construct a graph neural network model; Step 4 includes: Step 4.1: First, define the node and edge features; Step 4.2: A graph neural network model based on a supervised graph sparsification mechanism is adopted. For the infrastructure network structure of coastal cities, label information is introduced to guide edge sampling, constructing a task-related sparse subgraph. The model design includes the following steps: ① Edge probability learning: By introducing structural priors and label supervision information, the importance of each edge is learned; ② Sparse subgraph sampling: A sparse subgraph is generated based on the edge probability distribution, retaining the key structures that have the greatest impact on the evaluation results; ③ Graph neural network training: Node representation learning and resilience prediction are performed on the sparse graph structure. Step 5: After fusing the feature parameters of historical storm surge data with the coastal city infrastructure network structure data, input the data into the trained graph neural network model; based on the cascading impact propagation mechanism of the graph neural network model, simulate the step-by-step impact of storm surge disasters on the nodes of coastal city infrastructure, and obtain the dynamic changes of the temporal state of each node in the coastal city infrastructure network under the impact of storm surge. Step 6: Based on the graph neural network propagation results in Step 5, conduct a resilience assessment of the state evolution of each node in the coastal city infrastructure network under the influence of storm surge; based on the obtained resilience assessment index data of each node, summarize and analyze different functional areas to form an overall coastal city resilience assessment result.

2. The method for assessing the storm surge resilience of coastal cities based on graph neural networks according to claim 1, characterized in that: In step 1, the geographic information data includes high-resolution urban topography, geomorphological data, and coastline shape; infrastructure data includes port facility layout, seawall structural parameters, and residential area distribution data; and historical storm surge data includes the occurrence time, tide height, path trajectory, duration, wind speed and direction at landfall, and central pressure value of historical storm surges covering many years.

3. The method for assessing the storm surge resilience of coastal cities based on graph neural networks according to claim 1, characterized in that: In step 2, preprocessing includes outlier removal, missing value imputation, format unification, and standardization.

4. The method for assessing the storm surge resilience of coastal cities based on graph neural networks according to claim 1, characterized in that: In step 3, the entities of ports, seawalls, and residential areas are abstracted as nodes, and edge relationships are established based on their spatial proximity or functional association to construct a coastal city infrastructure network map, wherein: residential areas are connected by roads; edges are established between ports and the infrastructure that supports their transportation; and edges are established between seawalls and the areas protected by them.

5. The method for assessing the storm surge resilience of coastal cities based on graph neural networks according to claim 1, characterized in that: In step 4.1, a distinctive feature vector is defined for each type of node, including: residential area nodes: building density, total population, average building age, and shelter coverage; port nodes: annual throughput, cargo type, and facility disaster resistance level; seawall nodes: seawall length, height, structure type, service life, and maintenance cycle; edge features include: connection strength, connection reliability, and response time.

6. The method for assessing the storm surge resilience of coastal cities based on graph neural networks according to claim 1, characterized in that: In step 4.2, for edge probability learning: combined with node feature embedding, a multilayer perceptron is used to extract the correlation score of the edge.

7. The method for assessing the resilience of coastal cities to storm surge disasters based on graph neural networks according to claim 1, characterized in that: In step 4.2, for sparse subgraph sampling: to ensure the sampling process is supported by gradient optimization, Gumbel noise is introduced to achieve approximately discrete sampling. To construct a differentiable sparse edge sampling distribution, Softmax normalization is introduced, combined with a temperature annealing mechanism and temperature parameters. By controlling the smoothness of the distribution, a dynamic balance between exploration and determinism can be achieved.

8. The method for assessing the storm surge resilience of coastal cities based on graph neural networks according to claim 1, characterized in that: In step 4.2, for graph neural network training: the METIS graph partitioning algorithm is used to divide the original large graph into several local subgraphs to achieve distributed batch training; the overall training process is divided into three stages: teacher model training, student model distillation learning, and inference deployment.

9. The method for assessing the storm surge resilience of coastal cities based on graph neural networks according to claim 1, characterized in that: In step 6, the resilience assessment metrics for each node include: Node recovery time: refers to the time required for a node's functionality to recover from a damaged state to a normal state; Node function loss rate: refers to the proportion of node function that decreases during the impact of a disaster; Node connectivity interruption time: refers to the duration of interruption of communication or traffic functions between nodes due to cascading impacts.

Images