The Disaster AI Research Program at the Urban Resilience.AI Lab, Texas A&M University, develops artificial intelligence for disaster management and resilience. We build AI that reasons over geospatial data, infrastructure dependencies, and changing evidence so communities can anticipate disaster impacts, understand evolving conditions, and make better decisions under uncertainty. The program spans geospatial reasoning, disaster foundation models, disaster world models, agentic situational intelligence, cascading impacts and decision-flow AI, and flood and wildfire intelligence.
Disaster AI is the study and development of artificial intelligence systems that help communities anticipate disaster impacts, understand evolving conditions, and support better decisions under uncertainty. It brings together geospatial data, infrastructure dependencies, changing evidence, and human decision needs before, during, and after disasters.
Our lab studies Disaster AI as a scientific program rather than a single tool: how AI systems learn transferable representations of hazard and risk, reason over geospatial and infrastructure data, keep an auditable picture of conditions as evidence changes, and support human decision makers without hiding uncertainty or data limitations.
Disaster AI should move beyond isolated hazard maps, static dashboards, and one-off prediction tools. The next generation of disaster intelligence should reason over evolving community states, uncertain evidence, infrastructure dependencies, cascading impacts, and possible intervention choices.
Our lab develops the scientific foundations for that shift, and studies where these systems remain limited: sparse and delayed data, uneven coverage across communities, and the difficulty of validating models for safety-critical decisions.
We study foundation models and representation learning that encode disaster-relevant information about places, infrastructure systems, hazards, exposure, vulnerability, and recovery. These representations support downstream tasks such as flood impact prediction, wildfire risk assessment, road access loss, power outage anticipation, and critical-facility disruption.
Disasters affect communities through connected systems. Flooding can disrupt roads, road disruption can limit hospital or shelter access, and outages can affect medically vulnerable facilities. We develop context graphs, disaster ontologies, and geospatial reasoning methods that make these relationships explicit.
Disaster information arrives through sensors, official reports, imagery, news, local observations, and models. Those sources are incomplete, delayed, uneven, and sometimes contradictory. We develop disaster world models that track what is known, what is uncertain, what has changed, and which sources support or contradict a claim.
In disaster response, more information is not always better. The challenge is knowing which evidence is likely to change a decision, which source should be checked next, and when human review is needed. We study agentic information-seeking systems that optimize for decision value, provenance, and uncertainty awareness.
We study decision-flow AI: systems that organize information around the decisions communities actually face, such as resource prepositioning, facility protection, lifeline restoration, and public-warning updates, while keeping uncertainty, assumptions, constraints, and human agency visible.
Representations feed world models, world models ground agents, and agents support decision flows, with hazard-specific tools plugging in throughout.
Methods for reasoning over geospatial, tabular, network, and infrastructure data, including context graphs, disaster ontologies, query decomposition, and benchmarks that test traceable reasoning.
Shared representations of hazard, risk, resilience, exposure, vulnerability, and infrastructure state that transfer across places, hazards, and tasks.
Models of the changing state of community systems that connect baseline conditions with hazard signals, infrastructure status, facility access, lifeline disruption, and observed impacts.
Agents that seek, verify, and organize disaster evidence, reasoning about what was reported, where it applies, which sources agree or disagree, and whether the evidence is sufficient for a decision.
Methods that connect evidence to decision points, alternatives, constraints, uncertainty, and escalation needs rather than simply returning information.
Focused models for high-value tasks: flood depth and impact prediction, road and facility access assessment, wildfire risk and spread analysis, damage assessment, and lifeline disruption estimation.
Can AI answer complex disaster analytics questions with sound spatial logic and traceable evidence? We develop context graphs, disaster ontologies, and benchmarks that test whether AI systems reason correctly over geospatial data, infrastructure networks, hazard layers, and vulnerability indicators — measuring decision value, not just answer accuracy.
We study foundation models that learn transferable representations of hazard, risk, resilience, exposure, and infrastructure state — so disaster intelligence can transfer across places, hazards, and impact types instead of being rebuilt from scratch for every task and region.
Disaster information arrives as partial, delayed, and sometimes contradictory claims. We develop world models that maintain an evolving, auditable representation of community conditions — tracking sources, timestamps, confidence, and contradictions as new information emerges.
More information is not always better. We study AI agents that decide which evidence to seek, verify, or escalate next — optimizing for whether a new source would materially change a disaster decision, not for retrieval volume.
Disasters propagate through connected systems: flooding disrupts roads, road loss cuts hospital access, outages endanger vulnerable facilities. We model these cascading pathways and develop decision-flow AI that connects them to real decision points — resource prepositioning, facility protection, lifeline restoration — with assumptions, constraints, and human agency kept visible.
Flood and wildfire are our core application domains and the proving grounds for the broader program. Our work spans flood depth and impact prediction, road and facility access loss, lifeline disruption, wildfire risk and spread analysis, and damage assessment.
The full list is on the Disaster AI publications page.
3,000 rigorously verified questions across eight disaster types. A comprehensive test of whether language models can answer the questions disaster management actually asks.
Read “DisastQA: Benchmarking Question Answering in Disaster Management” on arXiv →Can LLMs orchestrate 26 disaster-response tools into executable multi-step workflows? A benchmark with step-level failure diagnosis across 14 models.
Read “DisasterBench: LLM Planning under Typed Tool Interface Constraints” on arXiv →A concept-to-schema knowledge graph that grounds AI reasoning over disaster data in expert knowledge and connects concepts to the geospatial schemas that answer them.
Read “DisasterLex: An Expert Knowledge Graph for Geospatial Disaster Analytics” on arXiv →Domain-aware coreset selection lets a tabular foundation model match supervised flood-depth accuracy using 0.7% of the training data and transfer to unseen watersheds.
Read “Data-Efficient Flood Depth Prediction with Tabular Foundation Models” on arXiv →Zero-shot damage classification from multi-view ground-level imagery, applied to the 2025 Eaton and Palisades fires, accelerating a critical recovery bottleneck.
Read “Automated Wildfire Damage Assessment via Vision-Language Models” on arXiv →A vision for collective human–machine intelligence for augmented resilience and the conceptual blueprint behind our agentic Disaster AI research program.
Read “Disaster Management in the Era of Agentic AI Systems” on arXiv →Parcel-level occupancy after Hurricane Helene, F1 0.848 — linking street-view imagery to recovery signals via vision-language models.
Read “Recov-Vision: Street View + VLMs for Post-Disaster Recovery” on arXiv →Instruction-tuned LLM for multi-label social-media text classification in disaster informatics.
Read “CrisisSense-LLM: Instruction Fine-Tuned Multi-label Classification” on arXiv →The Disaster AI program builds on more than a decade of the lab's research in urban resilience, including the Disaster City Digital Twin, human-infrastructure systems resilience to flooding, equity-aware infrastructure resilience assessment, and urban resilience to health emergencies. That foundation in how communities and infrastructure behave under stress grounds our AI research in real disaster dynamics.
We welcome collaborations with researchers, agencies, communities, and organizations working on disaster science, resilient infrastructure, geospatial intelligence, and human-centered AI. Prospective students and postdoctoral researchers can review opportunities to join the Urban Resilience.AI Lab.
Email Dr. Ali Mostafavi