ºÃÉ«ÏÈÉúTV

AI Decision Support for Rapid Post-Disaster Damage Assessment

Example of a sUAS used in response to Hurricane HeleneThe immediate aftermath of a disaster is a race against time where the first 72 hours are defined by a critical information gap. Emergency managers must rapidly determine where structures are most severely compromised and which transit routes remain passable for relief teams. Traditionally, this damage assessment has been an arduous manual process, requiring specialists to analyze imagery or conduct field surveys over several days. AI-SDM's research is actively closing this gap by integrating high-resolution small Uncrewed Aerial Systems (sUAS), crewed aircraft, and satellite imagery with cutting-edge computer vision and machine learning to transform raw pixels into operational intelligence.

While aerial imagery provides a wealth of data, its utility is often hindered by technical friction and the time lag of manual workflows. A primary technical hurdle identified by AI-SDM researchers is the variation in imagery resolution captured by these different sensors (sUAS, crewed aircraft, and satellites), as machine learning models trained on one source of imagery fail to transfer to others. During disasters, emergency managers do not know which imagery they will receive first to support their decisions, so ML models must be able to gracefully handle all resolutions.  To solve this, the Institute developed a suite of models focused on ML systems' ability to transfer across these differing resolutions and sources. Some of these models have been performant enough to be useful in practice, and these models—collectively known as the CLARKE (Computer vision and Learning for Analysis of Roads and Key Edifices) system—have been refined to run on a standard laptop in disconnected field environments. This effort was supported by the creation of the CRASAR-U-DROIDs dataset, the world’s largest open-source database of operationally relevant, parallel drone, crewed aircraft and satellite disaster imagery. This massive dataset includes 52 disaster scenes captured from 10 major disasters and features over 21,700 buildings, each independently labeled for damage, a feat achieved with the assistance of high school students in a unique citizen-science initiative.

An output of the building damage assessment model, indicating buildings with no damaged (green), minor damage (yellow), major damage (orange), or total damage (red)The core technical achievement lies in the deployment of Building and Road Damage Assessment (BDA/RDA) models capable of producing predictions in under 10 minutes. These algorithms classify building damage according to the Joint Damage Scale and road damage according to a custom scale developed with input from FEMA and the Florida Department of Emergency Management, providing a level utility that has not been deployed in the field with sUAS imagery before. Beyond detection, AI-SDM created path-planning algorithms that analyze road condition data to generate optimal post-disaster routes. During tabletop exercises with the Florida Division of Emergency Management (FDEM), these path visualizations were highly preferred by emergency managers for their ability to support immediate operational ingress and situational understanding. Furthermore, the research team is currently investigating aleatoric uncertainty——quantifying the inherent "noise" in model predictions—by comparing drone data with ground-truth FEMA assessments from Hurricane Ian to further refine model reliability; while also continuing the work of a CMU Capstone team to develop computer vision models to align pre-defined spatial data, like building polygons, with the imagery captured by the drones.

PhD student Tom Manzini explains the CLARKE System during a 2025 tabletop training exercise with disaster managersThe theoretical rigor of these models was proven through operational deployment during the 2024 hurricane season, including Hurricane/Tropical Storm Debby and Hurricane Helene. Working directly with the Florida State Emergency Response Team (SERT) and the Pennsylvania Emergency Management Agency (PEMA), AI-SDM delivered operationally valuable data products to emergency managers, assessing both building and road damage, that were then leveraged during their response. By automating the identification of damage that historically took specialists days to process, AI-SDM is fundamentally accelerating the transition from the response phase to long-term recovery.