🌍 Reasoning Over Earth Engine: An AI Agent for Kenya’s Climate Risk

Ask a question : “Which counties in Kenya face the highest flood risk this season, and who is most exposed?” , and get a grounded, ranked, explainable answer drawn from satellite data, population exposure, and authoritative Kenyan documents.

This project is an open replica of the Gemini-powered Geospatial Reasoning pattern announced at Google I/O 2026. It runs on Gemini 3.5 Flash and the free Earth Engine tier , it is not Google’s gated Geospatial Reasoning agent, but it demonstrates the same architecture, end to end, on Kenyan data.

---

Overview

A natural-language question is decomposed by a Gemini planner into a tool plan. An orchestrator runs the tools, each an expert “sub-agent”, and a fusion layer synthesizes a grounded answer:

• Knowledge Retrieval (RAG): retrieves Kenyan climate documents with sources
• Earth Engine: flood (Sentinel-1 SAR), rainfall (CHIRPS), elevation (SRTM)
• Population: exposed population per county (WorldPop)
• Risk Scoring: a transparent, auditable weighted score (no ML)

Everything degrades gracefully: with no API key or Earth Engine auth, the agent still runs using deterministic fallbacks, so it never hard-fails in a demo.

Architecture

                          Question
                             │
                             ▼
                   Gemini 3.5 Flash  (planner.py)
                             │   structured JSON plan
                             ▼
                   Orchestrator  (orchestrator.py)
                             │   dispatches tools (sub-agent pattern)
        ┌────────────────────┼────────────────────┬───────────────────┐
        ▼                    ▼                    ▼                   ▼
 search_knowledge   get_earth_engine_layers  get_population_    compute_risk_
   (RAG / FAISS)     (Sentinel-1/CHIRPS/SRTM)   exposure          score
        │                    │                    │                   │
        └────────────────────┴──────────┬─────────┴───────────────────┘
                                         ▼
                          Fusion  (fusion.py, Gemini 3.5 Flash)
                                         │  ranked + explained + cited
                                         ▼
                        Answer → Visualization (viz.py) → Streamlit UI

I/O 2026 concepts used

Gemini 3.5 Flash · agent orchestration · AGENTS.md / SKILL.md agent & skill definitions · tool-based reasoning · sub-agent pattern · grounded retrieval (RAG) · WebMCP-compatible tool design · Managed-Agents-style architecture.

---

Project structure

kenya-climate-agent/
├── src/
│   ├── agent/        planner.py · orchestrator.py · fusion.py
│   ├── tools/        earth_engine.py · population.py · risk_score.py · knowledge.py
│   ├── rag/          build_index.py · retriever.py
│   └── utils/        geo.py · viz.py