Step-by-Step Teaching Guide

STEP 1: The Core Problem — Why Regular AI Fails at Geospatial Tasks

The Challenge

Before understanding the solution, you need to understand why this problem is hard.

General-purpose AI models (like a standard chatbot) struggle with real-world spatial problems because:

Simple Analogy

Think of it like asking a general doctor to perform brain surgery. They understand medicine broadly, but lack the specialized tools, training, and step-by-step surgical protocols needed for that specific complex task.

---

STEP 2: The Solution — What Is a Geospatial Reasoning Agent?

Definition

A Geospatial Reasoning Agent is an AI system that combines:

• A powerful language model (Gemini)
• Specialized Earth science models
• Spatial data tools
• A structured planning process

...to answer complex questions about our planet that no single tool could answer alone.

Key Insight

The agent does not try to hold all the data in its memory at once. Instead, it:

User Question → Make a Plan → Select Right Tools → Execute in Workspace → Return Answer

This is called working outside the context window — processing large spatial data separately, then bringing results back.

---

STEP 3: The Four Key Capabilities (Building Blocks)

Think of these as four pillars holding up the agent's ability to function:

Pillar 1: 🔗 Connected Foundation Models

The agent connects to specialized AI models built specifically for Earth science:

It also connects to Google Maps Platform, Google Earth Engine, and Data Commons for additional data.

Pillar 2: 📋 Upfront Planning

Before doing anything, the agent:

1. Reads the user's question
2. Builds a step-by-step analytical plan
3. Identifies what data it needs
4. Selects backup datasets if local data is missing
5. Tracks where every piece of data came from

Why this matters: Without planning, the agent takes shortcuts — like using a web search instead of proper spatial tools — which produces unreliable answers.

Pillar 3: 📁 Private Data Integration

Organizations can bring their own data into the system:

• CSV spreadsheets (location data)
• GeoJSON files (custom map boundaries)
• Cloud database tables

The agent then joins private data with global Earth AI models — combining what the organization knows locally with what the AI knows globally.

Pillar 4: 🤖 Planetary Prediction Engine

This is the most advanced capability. Given local outcome data (like disease rates or crop yields), the agent:

1. Automatically selects relevant environmental and population signals
2. Fuses satellite and population data together
3. Trains a custom machine learning model for that specific prediction task

Real result: In public health research, this approach improved health outcome predictions by 12% and enabled mapping at 8 times finer spatial resolution.

---

STEP 4: How the Agent Works — The Workflow

Here is the complete process from question to answer:

┌─────────────────────────────────────────────────────┐
│  USER ASKS A QUESTION                               │
│  e.g., "Which neighborhoods face highest flood      │
│  risk given next week's forecast?"                  │
└──────────────────┬──────────────────────────────────┘
                   │
                   ▼
┌─────────────────────────────────────────────────────┐
│  PLANNING STAGE                                     │
│  • Break question into steps                        │
│  • Identify needed tools and datasets               │
│  • Plan fallbacks for missing data                  │
└──────────────────┬──────────────────────────────────┘
                   │
                   ▼
┌─────────────────────────────────────────────────────┐
│  TOOL SELECTION                                     │
│  • Pick relevant Earth AI models                    │
│  • Select appropriate datasets                      │
│  • Connect private user data if provided            │
└──────────────────┬──────────────────────────────────┘
                   │
                   ▼
┌─────────────────────────────────────────────────────┐
│  EXECUTION IN WORKSPACE                             │
│  • Process large spatial data outside context       │
│  • Run each step of the plan                        │
│  • Track all data sources                           │
└──────────────────┬──────────────────────────────────┘
                   │
                   ▼
┌─────────────────────────────────────────────────────┐
│  ANSWER DELIVERED                                   │
│  • Text explanation                                 │
│  • Tables of data                                   │
│  • Map visualizations                               │
│  • Full list of data sources used                   │
└─────────────────────────────────────────────────────┘

---

STEP 5: How Do We Know It Works? — The Evaluation

The Experiment Design

Researchers tested four versions of the agent to isolate what actually makes it better:

The Four Benchmarks Used

Two Key Findings

Finding 1: Domain-specific tools are essential

Adding Earth AI tools (Version 3 vs Version 2) produced a significant jump in performance on quantitative spatial tasks.

The one exception was UnivEARTH, where even the basic LLM scored 93.6% — because those answers already exist on the public web.

Finding 2: Planning is essential for complex tasks

For open-ended, multi-step problems (GeoExpertQueries), having tools alone was not enough. The agent needed upfront planning to stay on track and avoid shortcuts.

Important Limitation to Know

The GeoExpertQueries scores are measured by comparing the agent's process to expert-defined workflows — not by checking a single correct final answer. This means:

The agent produces a strong first draft for human experts to review — not a guaranteed final answer.

---

STEP 6: Real-World Applications

The agent has been tested across three major domains:

🏥 Public Health

• Mapping disease risk at neighborhood level
• Combining population movement data with environmental signals
• Predicting health outcomes with 12% better accuracy than demographic baselines alone

🌊 Disaster Response

• Flood forecasting combined with population density
• Road network routing to identify evacuation paths
• Crisis resilience planning

🏗️ Infrastructure Planning

• Site selection for facilities
• Solar potential mapping
• Grid outage prediction

---

STEP 7: Connecting the Concepts — Summary Framework

Use this framework to remember how everything fits together:

PROBLEM: Planetary data is too big, complex, and specialized for general AI

SOLUTION: Geospatial Reasoning Agent

    Built on 4 Pillars:
    ├── 1. Specialized Foundation Models (domain expertise)
    ├── 2. Upfront Planning (structured approach)
    ├── 3. Private Data Integration (real-world relevance)
    └── 4. Planetary Prediction Engine (custom ML training)

    Proven by 4 Benchmarks:
    ├── Tools alone help with narrow tasks
    └── Tools + Planning needed for complex tasks

    Applied to:
    ├── Public Health
    ├── Disaster Response
    └── Infrastructure Planning

---

Quick Knowledge Check

Test your understanding with these questions:

1. Why can't a standard LLM handle geospatial tasks well?
2. What happens during the planning stage before the agent executes a task?
3. What is the Planetary Prediction Engine and what does it produce?
4. Which benchmark showed that even a basic LLM performs well, and why?
5. Why is the agent's output described as a "first draft" rather than a final answer?

---

Key Takeaway: The Geospatial Reasoning Agent works not because it is simply a more powerful chatbot, but because it combines the right specialized tools with structured planning — the same way a skilled analyst would approach a complex problem methodically rather than guessing.