Residential Seismic Vulnerability Identification using Large Vision and Language Models

Abstract

One- and two-story wood-frame housing is the most common dwelling type in the United States. While it generally performs well in terms of earthquake life safety, certain seismic vulnerabilities, including crawlspace (cripple wall) foundations, living spaces over garages, and masonry chimneys, have led to high financial loss and displacement in past earthquakes. Building-level analyses have demonstrated significant performance differences due to foundation, garage, and chimney-related vulnerabilities, yet most regional seismic risk assessments use a single fragility function, selected based on era, to represent all wood-frame housing. Capturing the impact of specific vulnerabilities and targeted retrofit measures at the regional scale requires sufficiently detailed building inventory data. However, existing data rarely describes such features. This study addresses that gap by leveraging street-level imagery alongside Large Vision and Language Models (VLMs) to systematically collect relevant building features. Results show that VLMs can reliably screen images and identify broadly visible building features. Furthermore, this study contributes a reusable methodology for developing and evaluating prompts through repeatable strategies such as sequential questions to guide model behavior, explicit uncertainty accommodation, and assessment of uncertain or incorrect predictions, among others. While foundation types are inherently difficult to classify due to occlusions in the imagery and limitations of exterior building views, using probability-based outputs and leveraging neighborhood similarities enable nuanced assessments that could improve our understanding of exposure and uncertainty at the regional level.