Introduction
Disaster management has long been divided between pre-event planning and real-time observational monitoring. While tools like FEMA’s HAZUS provide detailed impact estimates before an event, and NASA’s mapping tools offer real-time visualization, a critical gap remains: the ability to quantify the causal drivers of disaster severity in real-time. This paper introduces a novel framework that leverages causal AI to move beyond simple prediction, enabling decision-makers to understand why a disaster is severe and simulate the impact of specific interventions.
Methodology: Causal AI for Disaster Assessment
The core of the proposed system is built upon three pillars of causal reasoning: Directed Acyclic Graphs (DAGs), Structural Causal Models (SCMs), and Causal Attribution.
Causal Modeling with DAGs
To capture the complex relationships between environmental factors and disaster impact, the researchers designed three distinct causal structures. These models range from assuming independent effects of features on severity to more complex structures where flooding acts as a root cause or infrastructure damage serves as a mediator for mobility and access disruption.
Attribution and Recourse
By training SCMs on these graphs, the system can perform "abduction, action, and prediction." This allows for the calculation of Necessity Scores, which identify the specific factors most responsible for a disaster's severity. Furthermore, the framework utilizes the DiCE (Diverse Counterfactual Explanations) framework to provide "algorithmic recourse"—suggesting the minimum set of changes (e.g., protecting specific infrastructure) required to reduce a county's severity classification from "High" to "Low."
Data Integration and Feature Extraction
The framework employs a multi-modal approach, combining "hard" physical data with "soft" human-centric data. The data pipeline processes three primary sources:
- Satellite Imagery: Using Google Earth Engine and land-cover maps from Dynamic World, the system captures 81 types of land-cover transitions (e.g., built-up area turning into water).
- News Articles & Social Media: LLMs (specifically GPT-4o) are used to extract mentions of damage to infrastructure, vegetation, and vehicles from Google News and Reddit.
- Ground Truth: Property damage data from the SHELDUS database serves as the target for training and validation.
Key Insights and Results
The study validated the framework using a dataset of "billion-dollar disasters" in the U.S. from 2017, covering 867 counties. Key findings include:
- Multi-Modal Synergy: Combining satellite data with social media and news reports provides a significantly more granular and accurate view of disaster impact than using any single source.
- Interventional Simulation: The tool allows users to toggle specific features—such as the area of built-up land turned into water—to see how those changes would affect the overall severity rating in real-time.
- Case Study Validation: In a test case of Hardin County, Texas (affected by 2017 tornadoes), the tool's attribution scores accurately identified that massive changes in "built-up area" and "bare land" were the primary drivers of the high severity rating.
Conclusion and Future Directions
This research demonstrates the potential of interventional causal models to transform disaster response from a reactive to a proactive process. By providing explainable attribution and actionable recourse, the tool empowers emergency responders to prioritize resources more effectively. Future work will focus on integrating live data streams for true real-time assessment and exploring causal discovery algorithms to automatically learn relationship structures from data.
Summary updated