Building vulnerability assessment method and system based on ai and GIS

Abstract

The application discloses a building vulnerability assessment method and system based on AI and GIS, a seismic risk analysis model is constructed by acquiring preset building seismic risk data and geographic space data set of a research area, a building damage index and a building vulnerability assessment index are acquired, and a vulnerability assessment model is constructed, a GeoIME ecosystem is constructed based on the seismic risk analysis model, the vulnerability assessment model and an LLM model, a GeoIME GPT system is constructed, vulnerability and risk assessment of building data of the research area are performed according to the GeoIME GPT system, and evaluation results are obtained, potential vulnerable buildings are identified by AI, global disaster resistance capability is improved, the GeoIME GPT system improves disaster management efficiency and realizes risk assessment automation, and a transparent reasoning path of the GeoIME-GPT enables users to trace back each analysis step, and uncertainty is converted from implicit risk into controllable variables.

Description

Technical Field

[0001] This invention belongs to the field of building vulnerability assessment technology, and particularly relates to a building vulnerability assessment method and system based on AI and GIS. Background Technology

[0002] Nearly two-thirds of the world's population will live in urban areas, presenting both significant challenges and opportunities for sustainable urban development. A key challenge lies in the urgent need to integrate artificial intelligence (AI) into urban systems to create smarter, safer, and more resilient livable cities. Driven by dynamic growth patterns, infrastructure demands, and environmental pressures, urban environments are becoming increasingly complex, requiring advanced AI solutions to support efficient urban planning. AI technologies that enable inter-system collaboration and ecosystem empowerment based on geospatial data, through the GeoOpenAI framework, may become crucial in shaping future cities. Without such integration, cities will face problems such as inefficient land use, disorderly expansion, unstable infrastructure, and inadequate services. Therefore, there is an urgent need to provide an AI and GIS-based method for building vulnerability assessment to address these technical challenges. Summary of the Invention

[0003] In view of this, the present invention provides a building vulnerability assessment method and system based on AI and GIS, which integrates artificial intelligence and geospatial intelligence for disaster management, risk assessment and global-scale mapping, enabling AI to identify potentially vulnerable buildings and improve global disaster resilience. The GeoIME GPT system has significant potential in improving disaster management efficiency, automating risk assessment and promoting AI-driven global urban resilience mapping, and is implemented using the following technical solutions.

[0004] In a first aspect, the present invention provides a building vulnerability assessment method based on AI and GIS, comprising the following steps: Obtain pre-defined building seismic risk data and geospatial dataset for the study area, and construct an earthquake risk analysis model based on the geospatial dataset and pre-defined building seismic risk data. The geospatial dataset includes seismic activity data, soil characteristic data, landslide susceptibility data, land use data, and data on proximity to dangerous areas. Obtain the building damage index and building vulnerability assessment index of the study area, and construct a vulnerability assessment model based on the building damage index and building vulnerability assessment index. The building vulnerability assessment index includes building structure, building material type, permeability, building quality and plot fine-grained parameters. The GeoIME ecosystem is constructed based on the earthquake risk analysis model, the vulnerability assessment model, and the LLM model. The GeoIME ecosystem is integrated with the GPT model to construct the GeoIME GPT system. The vulnerability and risk assessment of building data in the study area is performed based on the GeoIME GPT system to obtain the assessment results.

[0005] As a preferred embodiment of the above technical solution, the building damage index and building vulnerability assessment index of the study area are obtained, and a vulnerability assessment model is constructed based on the building damage index and building vulnerability assessment index, including: The expression for the building damage index is: (1) Wherein, DI is the damage index. It is the failure coefficient. It is the observed defect strength coefficient; It is the component effect coefficient. It is the connection effect coefficient. It is the basic effect coefficient. It is the isolator effect coefficient. It is the interaction effect coefficient, where i represents a natural number; A multi-level data collection strategy was adopted, and the core dataset of the GeoIME ecosystem was obtained by combining library research, field surveys and advanced spatial analysis. Based on the core dataset, geospatial feature layers are integrated, and vulnerability assessment and risk estimation are completed according to the FEMA-154 standard.

[0006] As a preferred embodiment of the above technical solution, a GeoIME ecosystem is constructed based on the earthquake risk analysis model, the vulnerability assessment model, and the LLM model, including: Data was collected and organized from multiple sources, including building structure data, land use maps, soil distribution maps, cadastral maps, earthquake and landslide maps, and satellite imagery; The collected datasets are integrated to adapt them to the GeoIME ecosystem. The adaptation process includes data cleaning, transformation and standardization, as well as geospatial data being geolocated, relocated to a common coordinate system and converted to a format compatible with GIS software and the GeoIME ecosystem. The standardized data will be used to identify correlations, spatial relationships and vulnerability patterns within the study area through spatial analysis, and the GeoIME ecosystem will be used to estimate the vulnerability and risk level of buildings. The GPT model is integrated into the GeoIME ecosystem to form the GeoIME GPT system, which is used for building safety hazard assessment and disaster risk prediction.

[0007] As a preferred embodiment of the above technical solution, the execution process of the GeoIME GPT system includes five stages: user query, query understanding, task planning and tool selection, geographic information execution, and output and decision support. The user query stage includes: Input: Accepts a single natural language query Q from the user; If provided, optional contextual metadata will be attached: region of interest (AOI) polygon, expected hazard type, building, time constraint, and execution preference, where expected hazard type includes landslide, earthquake, and soil type; Output: Original query Q and optional AOI, stored for source; User query representation: Representing a user's natural language query as follows: (2) in, It is a sequence of tokens representing the user's intent; Language Modeling (LLM) extracts entity information and relational attributes by parsing query Q. It then uses LLM to parse Q to extract semantic components for each subtask. The necessary natural language explanation for the subtask is generated, and the corresponding expression is: (3) (4) Where E represents a group of entities, Let E represent the individual entities in the set; m is the total number of entities in set E; and R represents the set of relationships between entities. Let k represent the individual relations in the set R, where k is the total number of relations in the set R.

[0008] As a preferred embodiment of the above technical solution, the query understanding execution process includes: The LLM model uses thought chain reasoning to decompose the query into subtasks T, the corresponding expression of which is: (5) Where G is the set of all available GeoIME operations, which includes proximity, buffer, classification, clustering, and risk analysis. LLM execution mapping: (6) in, Represents a function expressed by a fine-tuned LLM model; Internal operations: named entity recognition, dependency resolution, simple number extraction and reasoning representation, enabling the LLM model to emit an ordered list of subtasks; Output: A list T of subtasks with explanations.

[0009] As a preferred embodiment of the above technical solution, the execution process of task planning and tool selection includes: Each task Corresponding to a GeoIME operator The mapping is a definite table queried by the system, following formulas (7) and (8): (7) (8) Where X is the spatial feature set, and d represents the landslide buffer distance; This is the buffered spatial feature set, which contains all features whose distances to feature points in X are within the range d. X represents the input spatial feature set; D is the buffer distance parameter, indicating the range around each feature point in X that needs to be included in the buffer. Represents points in the original spatial feature set X. This represents any point on a two-dimensional spatial plane that may belong to the buffer zone. Point The Euclidean distance between x and x and Indicates that the condition ensures when When the distance to any point in X is within the range of d, it will be included in the buffer; Workflow is represented as an ordered sequence: (9) The LLM model is used to determine the execution order, and W is the complete workflow of the GeoIME task. It refers to a specific GIS operation in the workflow, representing the LLM parameter set integrated into GeoIME GPT; each This represents the learnable parameters in the LLM neural network architecture. The LLM model learns these parameters by training on a large text corpus, mapping a natural language input query x to a structured geospatial plan or instruction output. The LLM model processes text content expressions as follows: (10) Where z represents the interpretation, including task type, parameters, and intent; The output z will be passed to the GeoIME task planning phase for tool selection and geospatial execution. LLM will propose parameter values, which will default to domain-specific values ​​if not specified by the user. Output: Ordered workflow W and parameter set .

[0010] As a preferred embodiment of the above technical solution, geographic information execution includes: Building vulnerability assessment: for each building Each will be assigned a vulnerability score. In GeoIME, the workflow W is executed in a specified order, using the parameter g. Each tool call generates observation data and returns it to the LLM model, and records the operation log. Data integration: Import the external feature layers required by the W system, including building footprint and attributes, and disaster probability layers. The system utilizes satellite imagery and various auxiliary datasets; it automatically completes the preprocessing workflow, including coordinate reprojection, in-area object clipping, and topology verification; and it comprehensively evaluates each building using a weighted linear model to derive its vulnerability score. (11) in, It is a structural characteristic indicator. The weights assigned to each feature are based on the FEMA-154 score; Risk assessment: Applying disaster probability layers Combined with vulnerability, the risk index of each building is defined as follows: (12) in, Risk index for each building; Intermediate checks are performed after each tool call, including data consistency, attribute integrity, and value range. Failures will trigger an automatic rollback strategy or a manual alert. Output: marked and The building geospatial dataset, intermediate layers, and logs.

[0011] As a preferred embodiment of the above technical solution, the outputs and decision support include: Output generation: The final result can be generated in the following ways: Risk Map: Buildings are color-coded and have thematic overlays based on risk category. The corresponding expression is: (13) Reports and Recommendations: Summarize statistical data and decision-making rules, and generate FEMA-154 building repair reports; generate textual recommendations based on FEMA-154 guidelines, and provide data sources and reproducibility information; Output: Downloadable maps and reports; including LLM model extraction. , ; LLM recommends; ; GeoIME execution: ; For each remaining ,calculate and ; Output risk map and A list of buildings is provided for priority inspection.

[0012] Secondly, the present invention also provides a building vulnerability assessment system based on AI and GIS, applied to the aforementioned building vulnerability assessment method based on AI and GIS, comprising: The first model construction module acquires the pre-set building seismic risk data and geospatial dataset of the study area, and constructs an earthquake risk analysis model based on the geospatial dataset and the pre-set building seismic risk data. The geospatial dataset includes seismic activity data, soil characteristic data, landslide susceptibility data, land use data, and data on proximity to dangerous areas. The second model construction module obtains the building damage index and building vulnerability assessment indicators of the study area, and constructs a vulnerability assessment model based on the building damage index and building vulnerability assessment indicators. The building vulnerability assessment indicators include building structure, building material type, permeability, building quality and plot fine-grained parameters. An ecosystem construction module is used to build the GeoIME ecosystem based on the earthquake risk analysis model, the vulnerability assessment model, and the LLM model. The vulnerability assessment module is used to integrate the GeoIME ecosystem with the GPT model to construct the GeoIME GPT system, and to conduct vulnerability and risk assessments on the building data of the study area based on the GeoIME GPT system to obtain assessment results.

[0013] This invention provides a method and system for building vulnerability assessment based on AI and GIS. It acquires pre-defined seismic risk data and geospatial datasets of buildings in a study area, and constructs a seismic risk analysis model based on these datasets. This process yields building damage indices and vulnerability assessment indicators for the study area, and a vulnerability assessment model is built based on these indicators. A GeoIME ecosystem is constructed based on the seismic risk analysis model, vulnerability assessment model, and LLM model. The GeoIME ecosystem is then integrated with the GPT model to construct the GeoIME GPT system. The GeoIME GPT system is used to assess the vulnerability and risk of buildings in the study area, obtaining assessment results. This method integrates artificial intelligence and geospatial intelligence for disaster management and risk assessment, enabling AI to identify potentially vulnerable buildings and improving global disaster resilience. The GeoIME GPT system has significant potential in improving disaster management efficiency, automating risk assessment, and promoting AI-driven global urban resilience mapping. The transparent inference path of GeoIME-GPT allows users to trace each analysis step, transforming uncertainty from implicit risk into manageable variables. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A flowchart of the building vulnerability assessment method based on AI and GIS provided by this invention; Figure 2 The structural block diagram of the building vulnerability assessment system based on AI and GIS provided by this invention. Detailed Implementation

[0016] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0017] See Figure 1 This invention provides a building vulnerability assessment method based on AI and GIS, comprising the following steps: S1: Obtain the pre-defined building seismic risk data and geospatial dataset for the study area, and construct an earthquake risk analysis model based on the geospatial dataset and the pre-defined building seismic risk data. The geospatial dataset includes seismic activity data, soil characteristic data, landslide susceptibility data, land use data, and data on proximity to dangerous areas. S2: Obtain the building damage index and building vulnerability assessment index of the study area, and construct a vulnerability assessment model based on the building damage index and building vulnerability assessment index, wherein the building vulnerability assessment index includes building structure, building material type, permeability, building quality and plot fine-grained parameters; S3: Construct the GeoIME ecosystem based on the earthquake risk analysis model, the vulnerability assessment model, and the LLM model; S4: Integrate the GeoIME ecosystem with the GPT model to construct the GeoIME GPT system, and conduct vulnerability and risk assessments on the building data of the study area based on the GeoIME GPT system to obtain assessment results.

[0018] In this embodiment, the building damage index and building vulnerability assessment index of the study area are obtained, and a vulnerability assessment model is constructed based on the building damage index and building vulnerability assessment index, including: The expression for the building damage index is: (1) Wherein, DI is the damage index. It is the failure coefficient. It is the observed defect strength coefficient; It is the component effect coefficient. It is the connection effect coefficient. It is the basic effect coefficient. It is the isolator effect coefficient. It is the interaction effect coefficient, where i represents a natural number; A multi-level data collection strategy was adopted, and the core dataset of the GeoIME ecosystem was obtained by combining library research, field surveys and advanced spatial analysis. Based on the core dataset, geospatial feature layers are integrated, and vulnerability assessment and risk estimation are completed according to the FEMA-154 standard.

[0019] It should be noted that the GeoIME ecosystem is constructed based on the aforementioned earthquake risk analysis model, vulnerability assessment model, and LLM model. This includes: collecting and organizing data from multiple channels, including building structure data, land use maps, soil distribution maps, cadastral maps, earthquake and landslide maps, and satellite imagery; integrating the collected datasets to adapt them to the GeoIME ecosystem, wherein the adaptation process includes data cleaning, transformation, and standardization, as well as geospatial data being geolocated, retranslated to a common coordinate system, and converted to a format compatible with GIS software and the GeoIME ecosystem; using spatial analysis to identify correlations, spatial relationships, and vulnerability patterns within the study area using the standardized data, and using the GeoIME ecosystem to estimate the vulnerability and risk level of buildings; and integrating the GPT model into the GeoIME ecosystem to form the GeoIME GPT system, wherein the GeoIME GPT system is used for building safety hazard assessment and disaster risk prediction.

[0020] Breakthroughs in Large Language Models (LLMs) have opened new horizons for geospatial information management by combining natural language queries with professional Geographic Information System (GIS) operations. The GeoIME-GPT framework innovatively utilizes LLM as a semantic engine, autonomously parsing user queries, breaking them down into sub-tasks, and invoking relevant geographic information tools and data for spatial analysis and building assessment. This framework eliminates traditional manual processes, automatically completing building structure processing and analysis (e.g., extracting building types, exterior wall collapse risks, and adjacent relationships from cadastral maps, satellite imagery, or building inspector data). It also integrates geospatial data processing modules such as geological hazard risk assessment, liquefaction analysis, buffer zone delineation, and classification clustering, ultimately generating risk maps and analysis reports according to the FEMA-154 standard. By deeply integrating LLM into the GeoIME ecosystem, this framework leverages AI-driven reasoning capabilities to assist professionals in task planning and utilizes an adaptive language interface to implement a Building Risk Assessment and Early Warning System (RVS). For example, with GeoIME-GPT, users can interact with the system using natural prompts such as “Identify buildings with vulnerability >3 and interpret the map” or “Describe observable damage or degradation or other conditions that negatively impact the seismic performance of the building”. This invention proposes the GeoIME-GPT framework, which innovatively integrates artificial intelligence (AI) and geospatial intelligence (GPT) in disaster management, risk assessment, and global-scale mapping. This framework utilizes FEMA-154 guidance code to enable AI-based identification of potentially vulnerable buildings, contributing to improved global disaster resilience and supporting the achievement of Sustainable Development Goals (SDGs). This invention aims to explore the integration of open data, application programming interfaces (APIs), and GPT-based geospatial intelligence within the GeoIME platform to advance vulnerability assessment and risk prediction at both local and global levels. The performance of models from 87 buildings was analyzed through field surveys and GeoIME mapping. Furthermore, the results were cross-validated using regression tools in ArcGIS 10.8.3 to deepen the understanding of vulnerable buildings and disaster resilience. The GeoIME-GPT framework demonstrates significant potential in improving disaster management efficiency, automating risk assessment, and promoting AI-driven global urban resilience mapping.

[0021] It should be understood that by acquiring pre-defined building seismic risk data and geospatial datasets for the study area, and constructing a seismic risk analysis model based on the geospatial datasets and pre-defined building seismic risk data, the GeoIME GPT system is constructed. This involves obtaining building damage indices and building vulnerability assessment indicators for the study area, and then constructing a vulnerability assessment model based on these indicators. The GeoIME ecosystem is built upon the seismic risk analysis model, vulnerability assessment model, and LLM model. This GeoIME ecosystem is then integrated with the GPT model to construct the GeoIME GPT system. The GeoIME GPT system is used to assess the vulnerability and risk of building data in the study area, yielding assessment results. This system integrates artificial intelligence with geospatial intelligence for disaster management and risk assessment, enabling AI to identify potentially vulnerable buildings and improve global disaster resilience. The GeoIME GPT system has significant potential in improving disaster management efficiency, automating risk assessment, and promoting AI-driven global urban resilience mapping. The transparent reasoning path of GeoIME-GPT allows users to trace each analytical step, transforming uncertainty from implicit risk into manageable variables.

[0022] Optionally, the execution process of the GeoIME GPT system includes five stages: user query, query understanding, task planning and tool selection, geographic information execution, and output and decision support. The user query stage includes: Step 1, Input: Accept a single natural language query Q from the user (e.g., "Find vulnerable buildings with more than 3 floors"). If provided, optional contextual metadata will be attached: region of interest (AOI) polygon, expected hazard type, building, time constraint, and execution preference (e.g., fast or descriptive), where expected hazard type includes landslide, earthquake, and soil type; Output: Original query Q and optional AOI, stored for source; User query representation: Representing a user's natural language query as follows: (2) in, It is a sequence of tokens representing the user's intent; Language Modeling (LLM) extracts entity information and relational attributes (such as proximity, nearness, externality, and intersection) by parsing the query statement Q. It utilizes LLM to parse Q to extract semantic components for each subtask. (See step two), generate a natural language explanation of the necessity of the subtask (to improve transparency and for subsequent manual review), the corresponding expression is; (3) (4) Where E represents a group of entities, Let E represent the individual entities in the set; m is the total number of entities in set E; and R represents the set of relationships between entities. Let k represent the individual relations in the set R, where k is the total number of relations in the set R.

[0023] In this embodiment, the execution process of step two, query understanding (LLM semantic parsing + thought chain), includes: The LLM model uses thought chain reasoning to decompose the query into subtasks T, the corresponding expression of which is: (5) Where G is the set of all available GeoIME operations, which includes proximity, buffer, classification, clustering, and risk analysis. LLM execution mapping: (6) in, Represents a function expressed by a fine-tuned LLM model; Internal operations: Named entity recognition, dependency resolution, simple number extraction and reasoning representation, enabling the LLM model to emit an ordered list of subtasks instead of a single answer; Output: A list T of subtasks with explanations.

[0024] Specifically, step three, task planning and tool selection, includes the following steps: Each task Corresponding to a GeoIME operator The mapping is a definite table queried by the system, following formulas (7) and (8): (7) (8) Where X is a set of spatial features (e.g., buildings), and d represents the landslide buffer distance; This is the buffered spatial feature set, which contains all features whose distances to feature points in X are within the range d. X is the input spatial feature set; D is the buffer distance parameter, which represents the range around each feature point in X that needs to be included in the buffer (e.g., 500 meters around a building). Represents points in the original spatial feature set X. This represents any point on a two-dimensional spatial plane that may belong to the buffer zone. Point The Euclidean distance between x and x and Indicates that the condition ensures when When the distance to any point in X is within the range of d, it will be included in the buffer; Workflow is represented as an ordered sequence: (9) The LLM model is used to determine the execution order, and W is the complete workflow of the GeoIME task. It refers to a specific GIS operation in the workflow, representing the LLM parameter set integrated into GeoIME GPT; each This represents the learnable parameters in the LLM neural network architecture. The LLM model learns these parameters (weights or biases) by training on a large text corpus, mapping a natural language input query x to a structured geospatial plan or instruction output. For example, in the GeoIME-GPT system, if a user inputs a query (such as "find vulnerable buildings near the United Nations Headquarters in New York but not in a landslide zone"), the LLM model would process the text content as an expression like this: (10) Where z represents the interpretation, including task type, parameters, and intent; The output z will be passed to the GeoIME task planning phase for tool selection and geospatial execution. LLM will propose parameter values ​​(such as proximity or buffer distance d). If the user does not specify, the default value will be a domain-specific value. Output: Ordered workflow W and parameter set .

[0025] Specifically, step four, geographic information execution, includes: Building vulnerability assessment (FEMA-154): Each building is assessed according to the FEMA-154 guidelines. Each will be assigned a vulnerability score, and the workflow W will be executed in GeoIME in a specified order, using parameter g, where each tool call will generate observation data and return it to the LLM model (if configured), and record the operation log. Standard geoprocessing operators are defined in formula (8). Data integration: Import the external feature layers required by the W system, including building footprint and attributes (FEMA-154 field) and hazard probability layer. (Such as earthquake recurrence interval raster data), satellite impact data, and various auxiliary datasets; the system automatically completes the preprocessing process, including coordinate reprojection, in-area object clipping, and topology verification; a weighted linear model (i.e., a vulnerability scoring system inspired by the FEMA-154 standard) is used to comprehensively assess each building and derive its vulnerability score: (11) in, These are structural characteristic indicators (such as material, height, irregularity, age, and other parameters (FEMA-154)). The weights assigned to each feature are based on the FEMA-154 score; Risk assessment: Applying disaster probability layers Combined with vulnerability, the risk index of each building is defined as follows: (12) in, The risk index for each building (such as landslide occurrence, earthquake disaster, etc.); Intermediate checks are performed after each tool call, including data consistency (geometric validity), attribute integrity, and value range. Failures will trigger automatic rollback strategies (such as replacing the dataset) or manual alerts. Output: marked and The building geospatial dataset, intermediate layers, and logs.

[0026] Specifically, outputs and decision support include: Output generation: The final result can be generated in the following ways: Risk Map: Buildings are color-coded according to risk category (e.g., red: high, yellow: moderate, green: low) and subject overlays are applied (e.g., soil type, landslide area, seismic activity grid). The corresponding expression is: (13) Reports and Recommendations: Summarize statistical data and decision-making rules, and generate a FEMA-154 building repair report (including building ID, ...). , , (and categories); it also provides quantitative statistics for each risk category, spatial clustering indicators, and a recommended priority list.

[0027] The system also generates text recommendations based on FEMA-154 guidelines (e.g., "Inspect unreinforced masonry facades; prioritize..."). Key densely populated areas will be renovated, threshold It can be customized by the user or set according to policy (FEMA-154), providing data source and reproducibility information. The system also provides GeoIME tool version, dataset, LLM prompt template, parameter values, and timestamped log records. It generates text suggestions based on FEMA-154 guidelines and provides data source and reproducibility information; Output: Downloadable maps, reports (PDF / CSV / GeoJSON), and human-readable summaries, for example, given Q = "Find vulnerable buildings less than 3 stories near a certain building within an earthquake zone"; where the LLM model extracts... , ; LLM recommends; ; GeoIME execution: ; For each remaining ,calculate and ; Output risk map and A list of buildings is provided for priority inspection.

[0028] It should be noted that a multi-layered data collection strategy was adopted, combining library research, field surveys, and advanced spatial analysis to ensure depth and reliability.

[0029] (a) Library Research and Document Analysis: This phase primarily involved collecting secondary data from reliable sources, as well as relevant urban planning documents. Based on the FEMA-154 building code embedded in the GeoIME system, five core indicators for assessing building physical resilience were selected: building structure, building material type, permeability, building quality, and site-specific fine-grained parameters. Regression analysis tools (such as the Morans I index) were applied in the ArcGIS 10.8.3 environment. These indicators are widely recognized by the academic community as important parameters for assessing urban physical resilience. Furthermore, a literature review laid the theoretical and methodological foundation for subsequent analyses. (b) Field Surveys and Local Data Collection: The second phase involves large-scale field investigations in the city to validate and supplement secondary data. Field observations include building site surveys, community mapping, and records of urban fabric and construction quality; qualitative insights are gained regarding building construction techniques, building material standards, and building vulnerability. Furthermore, to support the building vulnerability and seismic risk assessment model built with GeoIME-GPT, (c) an open dataset integration scheme is adopted: integrating landslide data from the ArcGIS open data platform, information from OpenStreetMap and other open data sources, building footprint information, soil and elevation data, and land use data. These datasets are integrated through the GeoIME geodatabase API to ensure spatial reference consistency and interoperability for large-scale automated risk assessment.

[0030] (c) All collected data were systematically entered into the GeoIME geographic information system, forming the core dataset for the GeoIME-GPT analysis. By integrating geospatial element layers such as soil composition, landslide susceptibility, and building footprint, the system completed vulnerability assessment and risk estimation according to the FEMA-154 standard. The analysis framework integrated the GeoServer geographic service platform, database and GIS services, ArcGIS online platform, and Google API. These resources were deeply integrated with the GPT system through the GeoIME platform. This integrated design enabled automated spatial analysis, improved data interpretation accuracy, and efficiently completed the mapping of building vulnerability and seismic risk in the study area. Geospatial infrastructure, as a key area of ​​modern urban planning and infrastructure construction, is undergoing rapid development. With the increasing complexity of urban infrastructure systems, the application of geospatial intelligent technology has become an important means to improve the efficiency of building vulnerability assessment, planning and design, and management.

[0031] This invention employs a GeoIME-supported modeling framework to assess the vulnerability and risk of buildings before and after earthquakes. This framework integrates GIS with a FEMA-based building assessment model, constructing a comprehensive and flexible assessment system for analyzing structural vulnerability and potential damage. The framework is designed with scalability and flexibility in mind, adapting to various building types and sizes, and seamlessly integrating with new data sources, APIs, and modeling tools brought by new technologies. Data was first collected and organized from multiple sources, including building structure data, land use maps, soil distribution maps, cadastral maps, earthquake and landslide maps, and Google satellite imagery. Furthermore, empirical data on building design, human usage behavior, and surrounding environmental conditions were obtained through field investigations.

[0032] First, the collected datasets were integrated to adapt them to the GeoIME platform. This process included data cleaning, transformation, and standardization to ensure format consistency. Geospatial data was geolocated, unified to a common coordinate system, and converted to a format compatible with GIS software and the GeoIME platform. The standardized data underwent spatial analysis to identify correlations, spatial relationships, and vulnerability patterns within the study area. These analytical results were visualized through GIS layers in the ArcGIS and GeoIME interfaces, supporting interactive analysis and exploration. Subsequently, a building assessment model was developed using the preprocessed dataset to simulate structural performance before and after earthquakes. Employing the FEMA-154 code, the GeoIME-based model was used to estimate building vulnerability and risk levels. To enhance model reliability, spatial and structural parameters such as building design features, proximity effects, external fall hazards, occupancy differences, planar and vertical irregularities, water accumulation risk, and environmental conditions were incorporated into the simulation. The integrated model can generate recommended solutions, risk assessment maps, and reports for different building configurations. This method continuously monitors and provides feedback mechanisms to ensure that the GeoIME-enabled framework remains timely and accurate as new data and field information become available.

[0033] Ultimately, by integrating GPT into the GeoIME system, the GeoIME GPT system was formed. This integration provides building inspectors and analysts with automated auxiliary tools, enabling natural language interaction with the system. With the help of GeoIME's built-in GPT, users can query through simple commands (such as "write a hazard assessment report for a certain building and inform it of its vulnerability level" or "generate a risk map of high-rise buildings"). The GeoIME GPT model transforms the traditional manual operation process into an AI-based intelligent analysis process for building safety hazard assessment and disaster risk prediction. By integrating Large Model Language Understanding (LLM) into the geographic information engine, user natural language requests are transformed into an orderly GIS workflow, and vulnerability and risk assessment reports conforming to FEMA-154 standards are generated for specific building groups. The entire process consists of five stages: (1) user query; (2) query understanding (LLM); (3) task planning and tool selection; (4) geographic information execution (processing engine); and (5) output and decision support. Each stage is executed sequentially and may be iteratively optimized through manual verification.

[0034] To solve geospatial tasks, always use the following structure in your response: Thoughts: (Reflect on progress and decide the next step based on the user query and previous results; never skip this step); Actions: (Select the most suitable data and tools from {dataNames} and {toolNames} to solve the current step; explain the reasons for choosing the tool); Operational Inputs: (Provide specific input parameters for the selected tool based on FEMA-154 building structure data, required space data, vulnerability indicators, and user requirements); Observations: (Summarize the results of running the tool, such as vulnerability scores, risk categories, building classifications, and recommendations); (This loop may be repeated until the task is completely resolved, handling one step at a time). Or Thoughts: (Review the initial problem and reasoning process); Final Answer: (Output a complete solution to the original query, including risk assessment results, vulnerability classifications, and GeoIME-based explanations of recommendations).

[0035] Task Description: You are working with a dataset of 87 buildings in AB, Iran. Following the FEMA-154 guidelines, collect structural and non-structural information for each building. GeoIME-GPT must: interpret the user's requirements (e.g., "estimate building vulnerability scores," "map high-risk buildings," "generate a summary report"); select appropriate structural building features, geospatial data, and GeoIME tools (e.g., building type, soil classification, landslide classification, building classification, spatial connectivity, risk map); execute the tools sequentially to generate meaningful outputs (maps, scores, or reports), ensuring the outputs reflect FEMA-154 categories (e.g., risk score, vulnerability level). Throughout, always follow the format of "Thought → Action → Action Input → Observation" (or use "Final Answer" at the end); avoid tool illusions and only use tools explicitly defined in {dataStrings} and {toolStrings}; when using FEMA-154 building data, ensure that the input conforms to the building's attributes (structure type, height, age, materials, occupancy, proximity, irregularity, and other relevant data guided by FEMA-154); all subtasks must contribute to answering the original user's needs; the final output must clearly state the vulnerability classification (e.g., high risk, medium risk, low risk) and corresponding recommendations.

[0036] Question: Based on FEMA-154 attributes, determine which of the 87 buildings in Abhar (Iran) have a high risk of collapse under earthquake conditions, and generate a risk map; Idea: I need to evaluate a building dataset based on FEMA-154 attributes, classify vulnerabilities, and generate a risk map; Operation: ClassifyTool; Operation input parameters: "attributes" is "FEMA-154 standard data (covering 87 buildings)", and "criteria" is "structure type, building height, building year and number of users". Observation results: Each building has a vulnerability score (0-7); Idea: Now I should spatially map these scores to visualize the risk categories; Operation: RiskMappingTool; Input: {"scores": "Classification vulnerability score", "spatial data": "Building footprint"}; Observation results: A risk map was generated using building color coding (red = high, yellow = medium, green = low). Final conclusion: Of the 87 buildings, 48 ​​were classified as high-risk, 7 as medium-risk, and 32 as low-risk. The generated report and map highlight their spatial distribution. The prompts we used in our experiments were formatted specifically for GeoIME-GPT, enabling it to understand and process geospatial tasks and building structures. These prompts guided GeoIME-GPT to utilize various geospatial data, building structure features verified through field observations within the GeoIME database, and GIS tools to complete the corresponding risk assessment tasks.

[0037] To ensure the robustness of the GeoIME-GPT output, a multi-method spatial validation strategy based on mature geostatistical theory was adopted. This cross-validation scheme not only verifies the internal consistency of the vulnerability scores generated by the AI, but also aligns them with the classic spatial autocorrelation framework widely used in urban resilience research. To identify the pattern trends of vulnerable buildings categorized in the standard and urban housing resilience indicators, regression analysis tools were used in ArcGIS 10.8.3. The Morans I value tool was used to analyze the spatial distribution characteristics of vulnerable buildings and physical resilience in the urban housing structure of Ab, and the GeoIME GPT results were cross-validated. This analysis aimed to determine whether the distribution exhibited a clustered, random, or dispersed pattern; spatial weighting based on correlation was also used to improve the accuracy and reliability of the study. In the ArcGIS 10.8.3 environment, the spatial distribution characteristics of vulnerable buildings and the physical resilience of urban housing in Ab were revealed through the Morans I value spatial autocorrelation model. Furthermore, the Anselin local Morans I index method, clustering and outlier analysis tools, and other functional modules in the spatial statistics toolkit were used to study the spatial distribution characteristics of the physical resilience indicators of residents in Ab. To identify clusters of varying resilience values ​​in urban housing, the Gatiss-Oedgee (GI-Star) statistical model was used to analyze hot and cold zones, thereby providing a deeper understanding of the city's resilience distribution pattern. Spatial layer analysis of the final data revealed significant differences in resilience levels among different communities in Area A. This data is useful for assessing building vulnerability and the resilience of urban housing, and can serve as a reference for urban planning, crisis management, and promoting resilience against natural disasters.

[0038] Morans I is one of the most commonly used metrics for measuring and testing spatial autocorrelation. Compared with other spatial autocorrelation testing methods, Morans I exhibits superior statistical power when the model is misnomered. In recent years, this metric has been applied to texture measurement in object-oriented image classification. By analyzing the spectral and spatial features of adjacent pixels, the image is segmented into homogeneous regions. The calculation of Morans I uses the following formula (14): (14) Where n represents the total number of spatial phenomena observed through i and j. This represents the value of the i-th observation unit. This represents the average value across all observed units. It is binary The spatial weight matrix (i.e., the correlation between observation points i and j) is equal to the sum of the values ​​of all observation points; if This indicates the existence of a common boundary; This indicates that there is no common boundary, and the standard Z-test of the Moran statistic is calculated using formula (15): (15) in, and ...

[0039] Since the ordinary Moran index can only describe pattern characteristics, clustering and non-clustering analysis methods are used to study the spatial distribution patterns of map patterns. This tool can intuitively present the high or low value distribution characteristics of these phenomena within spatial clustering areas, as well as the effect values ​​that differ significantly from neighboring areas. Under the weighted complexity assumption, this analysis can identify clustering areas with similar complexity values ​​or similar sizes. This tool can effectively identify spatial non-clustering areas, defined in Moran's I Index, and its calculation formula is shown in equation (16): (16) in, These are attributes of complication i, and the average value of the corresponding attribute is given. , It is the space between complications i and j (Formula 17): (17) Where n equals the total number of geographical complications, and the standard score. See formula (18): (18) Here will be: (19) The positive Ii values ​​obtained by this equation indicate that high-value regions are surrounded by other high-value regions (high-high), while low-value regions are surrounded by similar regions (low-low); negative values ​​indicate that low-value regions are surrounded by high-value regions (low-high), or high-value regions are surrounded by low-value regions. Regional encirclement (high-low). The ability to identify statistically significant spatial clusters (HH, LL) and spatial outliers (HL, LH) is crucial for prioritizing intervention areas in urban earthquake risk management.

[0040] Specifically, the Getis-Ord Gi* technique in ArcGIS software is used to identify statistically significant high-value (hotspot) and low-value (coldspot) clusters. This is because the Getis-Ord Gi* statistic assesses the significance and strength of clusters through confidence surfaces and Z-scores. Positive Z-scores correspond to high-density clustering areas, with higher values ​​indicating more significant hotspots; conversely, negative values ​​reflect low-density clustering areas, with lower values ​​indicating more prominent colds. The formula for calculating the Getis-Ord JK statistic is as follows: (20) in, It is the attribute value of complication j. This represents the spatial weight between complications and i, where i, j, and n are the number of complications collected. (twenty one) (twenty two) Since Gi is a type of Z-score, it does not need to be recalculated.

[0041] Data on the physical disaster resistance of urban housing was preprocessed to ensure the robustness of the model, including data cleaning, transformation and division of real datasets and prediction datasets. To avoid overfitting and verify the generalization ability of the model, three independent statistical models need to be established: Moran's I index, local index and hot and cold zone model. The model verification adopts a dual-track mechanism: (1) internal statistical verification through geographic statistical error index; (2) external verification by domain experts. This hybrid verification framework ensures both the rigor of quantification and the contextual relevance of real disaster risk scenarios. Finally, the accuracy and performance of each model are evaluated using key indicators such as receiver operating characteristic (ROC) curve and root mean square error (RMSE). These evaluation results can intuitively reflect the model's ability to distinguish between vulnerable areas and disaster-resistant areas and its prediction accuracy. Root mean square error (RMSE): As one of the most commonly used statistical parameters in geographic information systems, RMSE has an important influence in geographic statistics applications. This invention uses this indicator to quantify the error difference between datasets, thereby accurately assessing the degree of deviation between the prediction model and the actual observation or real data. For the specific calculation formula, see equation (23): (twenty three) Where n is the number of data (samples), and is the sample size. The actual (target) value. It is the predicted value of sample i.

[0042] The performance of the classification model was evaluated using receiver operating characteristic (ROC) curves. The goal was to predict two different categories (e.g., "positive" and "negative"). The ROC curves show the trend of the error rate and success rate of the model's predictions as the decision threshold changes. The curves contain two axes: the vertical axis (True Positive Rate - TPR) represents the proportion of true positives, i.e., the percentage of samples correctly predicted as positive, calculated as shown in Equation 24. (twenty four) The horizontal axis (false positive rate - FPR) shows the false positive rate that is incorrectly identified as positive; in other words, it is the percentage of negative samples that are incorrectly predicted as positive, calculated using formula (25): (25).

[0043] The GeoIME GPT system successfully processed and analyzed a dataset of 87 buildings in AB City. By integrating FEMA-154 structural attributes and geospatial hazard data, it assessed the vulnerability and seismic risk of the buildings. Using the automated GeoIME GPT workflow, each building was assigned a vulnerability score. ) and disaster-weighted risk index ( The vulnerability map is presented visually in the GeoIME-GPT interface as a color-coded vulnerability map. A comprehensive evaluation of the GeoIME-GPT framework was conducted to examine its effectiveness in supporting geospatial professionals in conducting practical analysis tasks. This evaluation focused on the system's ability to interpret natural language queries, invoke applicable geospatial tools, and generate accurate and context-appropriate decision outputs, which will provide strong support for disaster risk prevention and urban resilience research.

[0044] To systematically evaluate the performance of GeoIME-GPT, 50 geospatial task queries were carefully designed to simulate real-world professional scenarios encountered by users and experts when constructing assessments. These query structures are similar to those in the Ab case study, adhering to three core criteria: (a) Professional Relevance—The query content reflects the actual analytical needs of GeoIME users, such as building assessment, recommendation formulation, and monitoring. For example, requests like "map high-risk earthquake buildings" or "estimate the probability of building collapse based on given data" realistically recreate professional workflows; (b) Tool Diversity and Workflow Complexity—To test the system's ability to handle complex geospatial reasoning, queries need to coordinate multiple analytical components such as spatial data extraction, geostatistical modeling, and visualization, achieved through GeoIME's internal toolchain; (c) Natural Language Understanding—The queries intentionally employ conversational questioning to evaluate GeoIME-GPT's ability to translate user natural language intent into structured geospatial operations and FEMA-154 instruction codes. This design ensures that the assessment truly reflects the natural interaction between end-users and the system.

[0045] However, by analyzing key elements in FEMA-154—including proximity of landslide zones to earthquake zones, building type, structural parameters, and natural disaster risks—geospatial analysis technology demonstrates unique advantages: it can accurately identify the most vulnerable buildings through the GeoIME-GPT system and achieve global risk assessment. This system not only effectively reduces the probability of building damage before or during earthquakes but also provides convenient query services for professionals and the public after disasters. For example, the platform's chat box displays the instant message, "Tell me about vulnerable buildings." <1> Users can interact with GeoIME through a language model to achieve deep integration of geospatial data and building information. The system can not only generate interactive maps but also output detailed reports, greatly improving communication efficiency between users and the platform. However, the language model occasionally provides incomplete answers, such as only providing building locations and vulnerability levels, or indicating that the task cannot be completed. Due to insufficient data, these issues of missing information and incorrect answers still require continuous improvement.

[0046] Furthermore, GeoIME-GPT analysis technology can effectively improve the timeliness of seismic maintenance and structural reinforcement of buildings, reduce downtime, and enhance the reliability of digital infrastructure. This technology can also be used to map disaster extent, identify critical infrastructure requiring repair or replacement, and assist users in writing prompts, generating automated reports, and facilitating user communication. This approach enables emergency responders to allocate resources and prioritize tasks more efficiently, significantly improving the speed and efficiency of disaster response and recovery efforts. The results summarized in Table 1 show that GeoIME-GPT achieved high accuracy in all 50 test scenarios. The framework driven by GPT-5-Turbo successfully solved 100% of the queries, while the GPT-4-Turbo model solved approximately 80%.

[0047] Table 1 Test Results

[0048] Table 1 shows the tool requirements statistics, indicating the number of different tools (such as GIS mapping, database retrieval, and vulnerability calculation) required by GeoIME-GPT. The required operation steps refer to the reasoning or execution steps needed to answer the queries (such as "loading data," "analyzing attributes," "generating maps," and "interpreting results"). The number of queries reflects the number of queries included in this complexity level (e.g., 12 simple queries using a single tool and 8 medium-complexity queries), while the number of correct responses (GPT-4 / GPT-5) refers to the number of queries correctly processed by each model (i.e., correct results, reasoning chains, and generated outputs verified through real data or expert review). The accuracy of the responses was measured by comparing the query performance of GPT-4 and GPT-5 under different numbers of tools / operation requirements. GPT-5-Turbo exhibited near-perfect performance (100% accuracy), demonstrating that advanced language models can reliably simulate expert-level geospatial reasoning capabilities when reasonably constrained by domain-specific rules (such as FEMA-154 logic) and tool enhancement architecture. This finding supports the hypothesis that when language models are integrated into a structured analysis environment, they will transcend their inherent limitations as "random parrots" and become trustworthy intelligent assistants in spatial decision-making.

[0049] The lower accuracy of the GPT-4-Turbo model primarily occurs when handling complex queries that require invoking five or more analytical tools or multi-step inference chains (e.g., multi-hazard vulnerability modeling or multi-raster dynamic classification). In contrast, GPT-5-Turbo, with its enhanced inference capabilities, contextual understanding, and spatial logic, significantly improves the ability to correctly select tools, perform GIS operations, and generate actionable insights. These results confirm the positive correlation between model performance and framework accuracy, demonstrating that using advanced language models can significantly improve the operational reliability of GeoIME-GPT. In particular, the integration of GPT-5-Turbo enables the system to autonomously complete multi-level geospatial analysis tasks, such as identifying high-risk structures, generating hazard maps, and producing comprehensive reports, with near-human accuracy and good interpretability.

[0050] Systematic evaluation demonstrates that GeoIME-GPT is an effective AI-assisted geospatial framework capable of bridging the gap between natural language interaction and advanced spatial analysis. Key findings include: (a) high accuracy and transparency: the framework provides a traceable reasoning process for each analytical step, ensuring interpretability and professional accountability; (b) strong adaptability: it successfully manages structured analytical workflows (such as FEMA-154 vulnerability assessment) and unstructured user queries; and (c) significant efficiency improvements: GeoIME-GPT automates complex multi-tool operations, reducing analysis time while maintaining methodological rigor, and is independent of model strength—the quality and inference depth of the underlying language model significantly impact accuracy and reliability. These results reinforce the framework's value as an intelligent, interpretable, and scalable solution, particularly suitable for geospatial AI-driven spatial decision-making in areas such as disaster risk reduction, infrastructure resilience, and Sustainable Development Goals monitoring.

[0051] Furthermore, structural analysis of selected buildings in AB City, conducted by extracting information from the GeoIME-integrated GPT model and combining ArcGIS tool analysis with field observations (cross-validation), revealed that 37% were metal frame structures, 35% were concrete frame structures, 9% were brick-concrete structures, and 18% had quality defects. These findings highlight the urgent need for authorities to prioritize strengthening urban physical resilience, ensuring accurate design and continuous structural monitoring (especially seismic performance) requires expert evaluation. Notably, the low-quality brick-concrete structures, accounting for 27% (approximately one-quarter) of the urban area, are mainly located in older communities and suburbs. The vulnerability of buildings in AB City is exacerbated by the widespread use of unstable building materials, which should be a key focus for the expert review panel. Of the 87 buildings selected as case studies, 63.218% belonged to the medium-to-high vulnerability category (i.e., low to medium resilience), while 36.781% were high resilience (mainly steel beams and brick-concrete structures). Field observations indicate that material instability remains a significant concern. Building quality is a key indicator for measuring the vulnerability and resilience of buildings; therefore, cities urgently need to strengthen disaster preparedness, especially their ability to respond to natural disasters and earthquakes. Through integrated GeoIME and GIS analysis, it was found that approximately 55.172% of residences in AB City exhibit high to very high vulnerability (i.e., low to very low disaster resistance). This assessment helps to more comprehensively analyze the spatial distribution and physical disaster resistance of vulnerable buildings in urban residential areas. The GeoIME-GPT system, based on LLM technology, was used to achieve efficient communication among users.

[0052] Evaluation results of the GeoIME-GPT framework demonstrate its significant potential in geospatial analysis, disaster risk prevention and control, and urban resilience planning. By integrating FEMA-154 structural attributes with geospatial disaster data, GeoIME-GPT successfully automated the seismic vulnerability assessment of 87 buildings in AB City. The generated vulnerability distribution map and disaster-weighted risk map clearly present the spatial distribution characteristics of high-risk buildings, providing a practical reference for professionals to prioritize structural reinforcement and risk prevention. This experiment verifies GeoIME-GPT's ability to bridge the gap between human reasoning and data-driven geospatial intelligence, pioneering a new paradigm for AI-assisted decision-making in spatial planning and risk management. A comparison of the evaluation results of the GPT-4-Turbo and GPT-5-Turbo language models reveals a significant correlation between model reasoning ability and analytical accuracy. While both can effectively handle query tasks of simple to moderate complexity, GPT-5-Turbo achieved full accuracy in all 50 test scenarios, including those requiring multi-tool collaboration and complex geostatistical inference. The contextual understanding depth and sequence reasoning capabilities of the large language model (LLM) directly determine the reliability of its generation of high-quality geospatial output.

[0053] See Figure 2 The present invention also provides a building vulnerability assessment system based on AI and GIS, applied to the aforementioned building vulnerability assessment method based on AI and GIS, comprising: The first model construction module acquires the pre-set building seismic risk data and geospatial dataset of the study area, and constructs an earthquake risk analysis model based on the geospatial dataset and the pre-set building seismic risk data. The geospatial dataset includes seismic activity data, soil characteristic data, landslide susceptibility data, land use data, and data on proximity to dangerous areas. The second model construction module obtains the building damage index and building vulnerability assessment indicators of the study area, and constructs a vulnerability assessment model based on the building damage index and building vulnerability assessment indicators. The building vulnerability assessment indicators include building structure, building material type, permeability, building quality and plot fine-grained parameters. An ecosystem construction module is used to build the GeoIME ecosystem based on the earthquake risk analysis model, the vulnerability assessment model, and the LLM model. The vulnerability assessment module is used to integrate the GeoIME ecosystem with the GPT model to construct the GeoIME GPT system, and to conduct vulnerability and risk assessments on the building data of the study area based on the GeoIME GPT system to obtain assessment results.

[0054] The three core contributions of this invention include: (a) developing a multivariate GIS dataset and a synchronous database server, which significantly improves the efficiency of geospatial information processing and cross-agency collaboration capabilities of governments at all levels in disaster risk prevention and control management; (b) constructing a national and global geospatial infrastructure ecosystem management system to automate building vulnerability assessment and risk prediction before and after earthquake disasters, while promoting the interconnection between users and the GeoIME-GPT platform; and (c) designing GeoIME-GPT workflows and supporting tools to accelerate the collaborative participation of professionals and the general public in disaster prevention, mitigation, and emergency management.

[0055] In all examples shown and described herein, any specific values ​​should be interpreted as merely exemplary and not as limitations; therefore, other examples of exemplary embodiments may have different values.

[0056] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0057] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A building vulnerability assessment method based on AI and GIS, characterized in that, Includes the following steps: Obtain pre-defined building seismic risk data and geospatial dataset for the study area, and construct an earthquake risk analysis model based on the geospatial dataset and pre-defined building seismic risk data. The geospatial dataset includes seismic activity data, soil characteristic data, landslide susceptibility data, land use data, and data on proximity to dangerous areas. Obtain the building damage index and building vulnerability assessment index of the study area, and construct a vulnerability assessment model based on the building damage index and building vulnerability assessment index. The building vulnerability assessment index includes building structure, building material type, permeability, building quality and plot fine-grained parameters. The GeoIME ecosystem is constructed based on the earthquake risk analysis model, the vulnerability assessment model, and the LLM model. The GeoIME ecosystem is integrated with the GPT model to construct the GeoIME GPT system. The vulnerability and risk assessment of building data in the study area is performed based on the GeoIME GPT system to obtain the assessment results.

2. The building vulnerability assessment method based on AI and GIS according to claim 1, characterized in that, Obtain the building damage index and building vulnerability assessment index for the study area, and construct a vulnerability assessment model based on the building damage index and building vulnerability assessment index, including: The expression for the building damage index is: （1） Wherein, DI is the damage index. It is the failure coefficient. It is the observed defect strength coefficient; It is the component effect coefficient. It is the connection effect coefficient. It is the basic effect coefficient. It is the isolator effect coefficient. It is the interaction effect coefficient, where i represents a natural number; A multi-level data collection strategy was adopted, and the core dataset of the GeoIME ecosystem was obtained by combining library research, field surveys and advanced spatial analysis. Based on the core dataset, geospatial feature layers are integrated, and vulnerability assessment and risk estimation are completed according to the FEMA-154 standard.

3. The building vulnerability assessment method based on AI and GIS according to claim 2, characterized in that, The GeoIME ecosystem is constructed based on the aforementioned earthquake risk analysis model, vulnerability assessment model, and LLM model, including: Data was collected and organized from multiple sources, including building structure data, land use maps, soil distribution maps, cadastral maps, earthquake and landslide maps, and satellite imagery; The collected datasets are integrated to adapt them to the GeoIME ecosystem. The adaptation process includes data cleaning, transformation and standardization, as well as geospatial data being geolocated, relocated to a common coordinate system and converted to a format compatible with GIS software and the GeoIME ecosystem. The standardized data will be used to identify correlations, spatial relationships and vulnerability patterns within the study area through spatial analysis, and the GeoIME ecosystem will be used to estimate the vulnerability and risk level of buildings. The GPT model is integrated into the GeoIME ecosystem to form the GeoIME GPT system, which is used for building safety hazard assessment and disaster risk prediction.

4. The building vulnerability assessment method based on AI and GIS according to claim 1, characterized in that, The execution process of the GeoIME GPT system includes five stages: user query, query understanding, task planning and tool selection, geographic information execution, and output and decision support. The user query stage includes: Input: Accepts a single natural language query Q from the user; If provided, optional contextual metadata will be attached: region of interest (AOI) polygon, expected hazard type, building, time constraint, and execution preference, where expected hazard type includes landslide, earthquake, and soil type; Output: Original query Q and optional AOI for storing the source; User query representation: Representing a user's natural language query as follows: （2） in, It is a sequence of tokens representing the user's intent; Language Modeling (LLM) extracts entity information and relational attributes by parsing query Q. It then uses LLM to parse Q to extract semantic components for each subtask. The necessary natural language explanation for the subtask is generated, and the corresponding expression is: （3） （4） Where E represents a group of entities, Let E represent the individual entities in the set; m is the total number of entities in set E; and R represents the set of relationships between entities. Let k represent the individual relations in the set R, where k is the total number of relations in the set R.

5. The building vulnerability assessment method based on AI and GIS according to claim 4, characterized in that, The query comprehension execution process includes: The LLM model uses thought chain reasoning to decompose the query into subtasks T, the corresponding expression of which is: （5） Where G is the set of all available GeoIME operations, which includes proximity, buffer, classification, clustering, and risk analysis. LLM execution mapping: （6） in, Represents a function expressed by a fine-tuned LLM model; Internal operations: Named entity recognition, dependency resolution, simple number extraction and reasoning representation, enabling the LLM model to emit an ordered list of subtasks; Output: A list T of subtasks with explanations.

6. The building vulnerability assessment method based on AI and GIS according to claim 5, characterized in that, The execution process of task planning and tool selection includes: Each task Corresponding to a GeoIME operator The mapping is a definite table queried by the system, following formulas (7) and (8): （7） （8） Where X is the spatial feature set, and d represents the landslide buffer distance; This is the buffered spatial feature set, which contains all features whose distances to feature points in X are within the range d. X represents the input spatial feature set; D is the buffer distance parameter, indicating the range around each feature point in X that needs to be included in the buffer. Represents points in the original spatial feature set X. This represents any point on a two-dimensional spatial plane that may belong to the buffer zone. Point The Euclidean distance between x and x and Indicates that the condition ensures when When the distance to any point in X is within the range of d, it will be included in the buffer; Workflow is represented as an ordered sequence: （9） The LLM model is used to determine the execution order, and W is the complete workflow of the GeoIME task. It refers to a specific GIS operation in the workflow, representing the LLM parameter set integrated into GeoIME GPT; each This represents the learnable parameters in the LLM neural network architecture. The LLM model learns these parameters by training on a large text corpus, mapping a natural language input query x to a structured geospatial plan or instruction output. The LLM model processes text content expressions as follows: （10） Where z represents the interpretation, including task type, parameters, and intent; The output z will be passed to the GeoIME task planning phase for tool selection and geospatial execution. LLM will propose parameter values, which will default to domain-specific values ​​if not specified by the user. Output: Ordered workflow W and parameter set .

7. The building vulnerability assessment method based on AI and GIS according to claim 6, characterized in that, Geographic information execution includes: Building vulnerability assessment: for each building Each will be assigned a vulnerability score. In GeoIME, the workflow W is executed in a specified order, using the parameter g. Each tool call generates observation data and returns it to the LLM model, and records the operation log. Data integration: Import the external feature layers required by the W system, including building footprint and attributes, and disaster probability layers. The system automatically completes preprocessing steps, including coordinate reprojection, in-area object clipping, and topology verification, using satellite imagery and various auxiliary datasets. A weighted linear model is then used to comprehensively assess each building and derive its vulnerability score. （11） in, It is a structural characteristic indicator. The weights assigned to each feature are based on the FEMA-154 score; Risk assessment: Applying disaster probability layers Combined with vulnerability, the risk index of each building is defined as follows: （12） in, Risk index for each building; Intermediate checks are performed after each tool call, including data consistency, attribute integrity, and value range. Failures will trigger an automatic rollback strategy or a manual alert. Output: marked and The building geospatial dataset, intermediate layers, and logs.

8. The building vulnerability assessment method based on AI and GIS according to claim 7, characterized in that, Outputs and decision support include: Output generation: The final result can be generated in the following ways: Risk Map: Buildings are color-coded and have thematic overlays based on risk category. The corresponding expression is: （13） Reports and Recommendations: Summarize statistical data and decision-making rules, and generate FEMA-154 building repair reports; generate textual recommendations based on FEMA-154 guidelines, and provide data sources and reproducibility information; Output: Downloadable maps and reports; including LLM model extraction. , ; LLM recommends; ； GeoIME execution: ; For each remaining ,calculate and ; Output risk map and A list of buildings is provided for priority inspection.

9. A building vulnerability assessment system based on AI and GIS, characterized in that, The method for assessing building vulnerability based on AI and GIS as described in any one of claims 1-8 includes: The first model construction module acquires the pre-set building seismic risk data and geospatial dataset of the study area, and constructs an earthquake risk analysis model based on the geospatial dataset and the pre-set building seismic risk data. The geospatial dataset includes seismic activity data, soil characteristic data, landslide susceptibility data, land use data, and data on proximity to dangerous areas. The second model construction module obtains the building damage index and building vulnerability assessment indicators of the study area, and constructs a vulnerability assessment model based on the building damage index and building vulnerability assessment indicators. The building vulnerability assessment indicators include building structure, building material type, permeability, building quality and plot fine-grained parameters. An ecosystem construction module is used to build the GeoIME ecosystem based on the earthquake risk analysis model, the vulnerability assessment model, and the LLM model. The vulnerability assessment module is used to integrate the GeoIME ecosystem with the GPT model to construct the GeoIME GPT system, and to conduct vulnerability and risk assessments on the building data of the study area based on the GeoIME GPT system to obtain assessment results.

Images