OSGR: Observation-Support Graph Reasoning for Earth Observation Question Answering

| Source: arXiv AI

Tags: Earth Observation, OSGR, GPT-4o, Earth-Bench, DisasterM3-Reason, Geospatial AI

OSGR, a two-stage graph reasoning framework for Earth observation question answering, lifts GPT-4o accuracy from 49.3% to 63.7% on Earth-Bench and DisasterM3-Reason while cutting execution steps.

Details

Answering geographic questions over Earth observation archives requires finding the right observation items and deriving results under strict space and time constraints. The authors say current LLM agents either fail to ground items or face search-space explosion. OSGR splits the work into offline symbolic concept deduction over an Observation-Support Graph and online Relation-Conditioned Grounding that uses spatiotemporal compatibility kernels to pick grounded observation sets. On Earth-Bench and DisasterM3-Reason, GPT-4o accuracy rose from 49.3% to 63.7% and the execution-step ratio fell from 2.9 to 2.5. This matters for geospatial and disaster-response teams building agentic analysis on satellite archives.