Research Experiences for Undergraduates (REU), NSF NHERI 2025: Evaluating Vision-Language Models for Automated Wildfire Damage Classification Using Street-View Imagery

This project aims to improve the efficiency of post-wildfire responses by investigating the potential of artificially intelligent Vision Language Models (VLMs) in automating wildfire damage assessment. Ground-level imagery documenting a street-view perspective of residential structures impacted by the 2025 Los Angeles Wildfires was collected through the RAPID facility and used for classification and testing. A project parallel to this one was simultaneously conducted by Ryan Chen but using aerial imagery rather than street-level. Employing a granular damage classification index, called the Combustion Hierarchy Scale (CHS Index), the capabilities of two VLMs: 1) Gemini and 2) Claude.ai were tested against a sample ground truth dataset that was curated through manual classification. Results from VLM testing were compared and mapped using ArcGIS Pro to ultimately create a sample 3D Damage Map indicating the distribution and severity of wildfire damage within neighborhoods in Los Angeles, CA. Most notably, several materials used within the study are adaptable for other research. The panoramic images used for VLM testing are a part of a repository of over 200,000 street-level imagery located in DesignSafe that are openly accessible to the research community for future exploration. In addition, an initial multi-class damage index and prompt were developed for testing and shown to be effective in assessing damage severity via pre-trained vision language models. Past research in this domain has been early applications of artificial intelligence and binary classifiers. However, binary classifications ultimately fail to capture the granular spectrum of damage, which the Combustion Hierarchy Scale, used in this study, can assess. Furthermore, artificial intelligence, specifically vision language models are a recent development, having only emerged within the past five years. The integration of both these advancements provides key information for recovery responses (e.g. distribution of first defenders, allocation of aid, etc.) and post-disaster investigations focused on burn intensity, urban fire behavior, and environmental contaminants. Ultimately, the contributions of this study consist of 1) an openly available imagery dataset, 2) a preliminary damage severity index and VLM prompt, and 3) a framework for integrating pre-trained VLMs into advancing disaster response strategies and long-term research in community resilience.