Disaster-GeoRAG

Inspiration

When disasters strike, responders need fast, reliable insights from satellite imagery. Current tools are often either too generic, too slow, or prone to hallucination. We wanted to create a system that can answer disaster questions grounded in authoritative knowledge rather than guess.

That inspired Disaster-RAG: a Retrieval-Augmented Generation (RAG) system for disaster triage on Earth observation imagery.

What it does

Disaster-RAG takes an EO image and optional metadata (sensor, date, cloud cover), retrieves relevant cues from a curated disaster knowledge base, and generates structured outputs with explainability traits.

Inputs: Satellite image + metadata
Retrieval: Top-k disaster cues from Copernicus EMS & NASA Earthdata
Outputs:
- is_disaster → true/false/uncertain
- type → one of 8 disaster categories (flood, wildfire, storm/wind, earthquake, landslide, volcano, drought, industrial)
- confidence score
- rationale citing retrieved evidence

We also built a simple Gradio app with two modes:

Triage: “Is this a disaster? If yes, what kind?”
Ask: free-form Q&A, with retrieved evidence shown alongside the answer.

How we built it

Dataset: GAIA (28.9k train, 3.9k val, 8.17k test) as our testbed for evaluation.
Knowledge Base: 12 short disaster cues distilled from Copernicus EMS & NASA Earthdata. Each cue has an ID, type, short description, and source link. Embedded with Sentence-Transformers, indexed in FAISS.
RAG Pipeline:
1. Extract image metadata.
2. Retrieve top-k cues from KB.
3. Feed image + metadata + cues into the VLM.
4. Generate structured JSON output.
Interface: Built with Gradio for quick demo (Triage & Ask tabs).

Challenges we ran into

Curating a compact but authoritative disaster knowledge base.
Designing prompts that enforce structured JSON output.
Handling uncertainty: encouraging the system to say “uncertain” when evidence is weak.
Building an end-to-end pipeline in limited time.

Accomplishments that we're proud of

Built a working Disaster-RAG pipeline in under 8 hours.
Designed a compact disaster taxonomy (8 categories) with grounded cues.
Created a Graphic User Interface (GUI) that shows not only answers, but also retrieved grounded evidence, boosting transparency.
Established a foundation for explainable, uncertainty-aware disaster triage.

What we learned

RAG improves trustworthiness by grounding answers in authoritative sources.
Structured outputs make AI predictions easier to evaluate and integrate downstream.
Transparent evidence citations can shift models from “black boxes” to decision-support tools.
How to quickly adapt large models to EO disaster tasks without heavy fine-tuning.