GeoDisaster benchmark tests AI agents with 2,921 disaster scenarios
Multi-agent framework outperforms existing VLMs on 43 question types...
Deep Dive
Researchers introduce GeoDisaster, a benchmark of 2,921 verified instances across 5 disaster task families (deforestation, multi-hazard, building damage, flood routing, SAR flood monitoring). It uses an orchestrated multi-agent framework with 18 disaster-oriented tools, coordinated via Role-Contract Expectation Alignment (RCEA). The framework challenges current RS-VLMs and improves tool use, evidence grounding, and decision generation for operational geo-intelligence.
Key Points
- GeoDisaster includes 2,921 verified instances across 43 question types and 5 disaster task families
- Framework uses 18 disaster-oriented tools with role-specialized agents coordinated via execution contracts
- RCEA training improves tool use, evidence grounding, and decision generation over existing RS-VLMs
Why It Matters
Accelerates AI-powered disaster response with verifiable, tool-grounded intelligence from satellite data.