From Local Data to Global Risk Maps: A Geospatial Generative Pre-Trained Transformer Framework for Earthquake Building Vulnerability Estimation

Earthquake vulnerability assessment of buildings within geographical information systems (GIS) remains a complex and fragmented process, often requiring the integration of multiple analytical methods and heavy reliance on expert interpretation. Despite advances in geospatial technologies, existing workflows for vulnerability assessment and reporting are still labor-intensive, loosely connected, and difficult to automate. Recent developments in Artificial Intelligence (AI), particularly large language models (LLMs) such as Chat Generative Pre-trained Transformer (GPT) and generative frameworks like AutoGPT, offer new opportunities to streamline and automate complex geospatial processes through semantic reasoning and adaptive task execution. The objective of this work is to introduce a framework and implementation of the Geospatial Infrastructure Management Ecosystem (GeoIME), powered by the geospatial GPT. The proposed GeoIME-GPT approach integrates AI-driven reasoning with GIS-based analysis to assist building inspectors and decision-makers in automating map generation, performing vulnerability and risk assessments, and generating rehabilitation recommendations in accordance with Federal Emergency Management Agency (FEMA) 154 guidelines and geospatial data analytics. Among these, an incorporated LLM understands natural language questions, breaks them down into geospatial subtasks, and eventually calls the necessary GeoIME tools for spatial analysis and building evaluation in sequence. The framework ’learns’ as it progresses and, depending on the context, becomes more or less precise, leading to an intuitive language-based interaction with geospatial data. From an evaluation of 87 buildings through field observations, structural data extraction, and GIS-based risk mapping, together with 50 geospatial task queries, more efficient, accurate, and interpretable assistance for building inspectors was demonstrated compared to Rapid Visual Screening (RVS). By employing two LLMs (GPT-4 (80%) and GPT-5 (94%), the system achieves high accuracy, and by optimizing spatially, a further increase in analysis accuracy (+12%) is obtained. Despite some data privacy and algorithm optimization issues, these results demonstrate GeoIME-GPT’s promising capability in automated seismic risk estimation and building vulnerability assessment, in accordance with FEMA-154 (i.e., Building Seismic Safety Council) standards, as well as for sustainable buildings.