After studying this material, students should be able to:

1. Explain what the Planetary Prediction Engine (PPE) is and the problem it solves
2. Describe the three-stage modular workflow of PPE
3. Understand why multimodal fusion outperforms single-modality approaches
4. Evaluate PPE's real-world performance across different domains
5. Articulate the broader significance of autonomous geospatial AI for humanitarian applications

---

Step-by-Step Teaching

Step 1: Understanding the Problem First

Before appreciating PPE, you need to understand why it was needed.

The Core Challenge

Humanity faces urgent global problems:

• 🦠 Disease outbreak prediction
• 🌾 Food security forecasting
• 🌍 Environmental disaster risk
• 📊 Socioeconomic vulnerability mapping

All of these require geospatial modeling — building mathematical models that predict outcomes across geographic locations.

Why Was This Hard Before PPE?

💡 Key Insight: Existing AI tools (AutoML, LLM agents) worked well for standard data but failed at geospatial workflows because they assumed data was already clean and organized.

---

Step 2: What Is the Planetary Prediction Engine (PPE)?

Simple Definition

PPE is an autonomous AI system that:

• Takes a plain English question about the world
• Automatically finds, cleans, and organizes geospatial data
• Trains a predictive model
• Delivers results and a report

Before vs. After PPE

BEFORE PPE:
Natural language question
        ↓
    [Weeks of manual work by expert teams]
        ↓
    Predictive model

AFTER PPE:
Natural language question
        ↓
    [Minutes of autonomous AI processing]
        ↓
    Predictive model + Report

Where Does PPE Fit?

PPE is part of Google Earth AI — a broader initiative turning planetary-scale information into actionable insights.

---

Step 3: How PPE Works — The Three-Stage Workflow

PPE breaks the entire modeling process into three independent stages, each managed by a Large Language Model (LLM).

🔵 Stage 1: Intelligent Data Selection

What happens here:

• The LLM reads your natural language query
• It searches across planetary-scale data sources
• It selects the most relevant datasets for your prediction task

Example: If you ask "Where is food insecurity highest in Nigeria?", the system automatically identifies that market prices, rainfall data, and crop yield data are relevant.

Key design choice: Data is passed between stages as opaque handles (like file pointers), NOT copied into the LLM's memory.

💡 Why opaque handles matter: LLMs have limited "context windows" (how much text they can process at once). Passing actual data would overflow this limit. Handles are just references — like a library card number instead of the whole book.

---

🟡 Stage 2: Dataset Curation

What happens here:

• Raw data is cleaned and preprocessed
• Features are engineered (transformed into useful inputs for models)
• Spatial validation is performed — ensuring geographic data aligns correctly

This stage is where PPE truly differs from standard AutoML, because geospatial data has unique challenges:

• Different coordinate systems
• Different spatial resolutions
• Temporal misalignment across datasets

---

🟢 Stage 3: AutoML & Prediction

What happens here:

• Models are automatically trained and tested
• Best-performing model is selected
• A comprehensive report is generated with results

Output: Predictions + explanations, ready for researchers or policymakers.

---

Step 4: The Secret Weapon — Multimodal Fusion

This is one of the most important technical concepts in the article.

Two Types of Data PPE Combines

Why Combine Both?

Think of it like diagnosing a patient:

• Statistical covariates = blood test results (explicit, measurable)
• Embeddings = doctor's intuition from years of experience (complex pattern recognition)

Neither alone is as powerful as both together.

Statistical Covariates    +    Foundation Model Embeddings
(explicit signals)              (latent complex patterns)
         \                              /
          \                            /
           ——————> MULTIMODAL FUSION <——————
                         |
                  Better Predictions

💡 Ablation studies (tests where you remove one component at a time) confirmed this — removing either modality consistently made performance worse.

---

Step 5: Real-World Performance — Did It Actually Work?

PPE was tested across three very different real-world scenarios:

🏥 Case 1: US Public Health (CDC Health Indicators)

• Task: Predict 21 different health indicators across US regions
• PPE Result: Mean R² = 76.8%
• Manual Expert Baseline: Mean R² = 60.0%
• Improvement: +16.8 percentage points

📊 R² explained: R² (R-squared) measures how well a model explains variation in the data. 100% = perfect prediction. 0% = no better than guessing the average. Higher is better.

---

🌾 Case 2: Food Security Downscaling in Nigeria

• Task: Take food insecurity data at the state level (ADM1) and predict it at the finer local government area level (ADM2)
• Why hard: Less data available at local level; local conditions vary dramatically
• PPE Result: R² = 66.1%
• Baseline: R² = 31.5%
• Improvement: PPE doubled baseline accuracy

How PPE achieved this: By autonomously integrating:

• Local market price shocks
• Food price anomalies
• Microclimate indicators

---

🦠 Case 3: Ebola Outbreak Nowcasting in DRC (2026)

• Task: Predict which health zones would be newly invaded by Bundibugyo ebolavirus each week
• Metric used: Recall@10 — "Of the top 10 predicted zones, how many actually became outbreak zones?"
• PPE Result: 83.3% (correctly identified 15 of 18 newly invaded zones)
• State-of-the-art Bayesian baseline: ~73%
• Improvement: +10.3 percentage points

💡 Why Recall matters here more than accuracy: In outbreak prediction, missing a real outbreak zone is catastrophic. You want to catch as many true cases as possible, even if it means some false alarms.

---

Step 6: Connecting the Dots — Why This Matters

The Democratization Argument

Before PPE:

Research Question → Specialized Engineering Team → Weeks → Model

After PPE:

Research Question → PPE → Minutes → Model

This means:

• Humanitarian organizations without large technical teams can now build models
• Policymakers can get answers faster during crises
• Researchers can focus on asking better questions instead of engineering data pipelines

The Shift in Human Role

---

Step 7: Limitations and Future Directions

PPE is described as early-stage research, not a finished product. Important honest caveats:

• Currently limited in which geospatial data sources it can access
• Foundation model embeddings used are still expanding (e.g., Remote Sensing Foundations embeddings planned)
• Results need continued validation across more geographies and domains

Future work includes:

• More geospatial data sources
• Additional foundation model embeddings
• More use case exploration

---

Summary: The Big Picture

PROBLEM: Geospatial modeling is slow, manual, and requires expert teams

SOLUTION: PPE — autonomous 3-stage AI workflow
    Stage 1: Intelligent Data Selection (LLM finds relevant data)
    Stage 2: Dataset Curation (automated cleaning & feature engineering)
    Stage 3: AutoML & Prediction (model training + report)

KEY INNOVATION: Multimodal Fusion
    Statistical covariates + Foundation model embeddings = better predictions

RESULTS:
    ✅ +16.8% improvement in US health indicators
    ✅ 2x improvement in Nigeria food security downscaling
    ✅ +10.3% improvement in DRC Ebola outbreak prediction

IMPACT: Democratizes planetary-scale geospatial modeling
    → Faster response during humanitarian crises
    → Accessible without specialized engineering teams

---

Quick Self-Check Questions

1. What problem does PPE solve that existing AutoML tools could not?
2. Why are data artifacts passed as "opaque handles" between stages?
3. What is the difference between statistical covariates and foundation model embeddings?
4. Why is Recall a more appropriate metric than accuracy for Ebola outbreak prediction?
5. In your own words, how does PPE "democratize" geospatial modeling?