Session 3: Emerging AI Trends in Earth Observation

Foundation Models, Self-Supervised Learning, and Explainable AI

Instructor

Stylianos Kotsopoulos

Date

November 17, 2025

Session Overview

This 2-hour session introduces three revolutionary AI trends reshaping Earth observation: Geospatial Foundation Models (pre-trained on massive satellite archives), Self-Supervised Learning (learning from unlabeled data), and Explainable AI (understanding model decisions). These technologies address key challenges in EO: limited labeled data, model interpretability, and generalization across diverse regions.

NoteLearning Objectives

By the end of this session, you will be able to:

  1. Explain what foundation models are and why they’re transformative for EO
  2. Identify major geospatial foundation models (Prithvi, Clay, SatMAE, DOFA)
  3. Understand self-supervised learning approaches for unlabeled satellite data
  4. Apply explainable AI techniques (SHAP, LIME, Grad-CAM) to interpret models
  5. Assess how these trends address Philippine EO challenges
  6. Evaluate when to use foundation models vs. training from scratch

Presentation Slides

Session Details

Duration: 2 hours (120 minutes) | Format: Theory + Demonstrations | Difficulty: Advanced

Prerequisites: - Understanding of CNNs and LSTMs (Days 2-4) - Transfer learning concepts - Basic model training experience - Familiarity with overfitting and generalization

Materials Provided: - Presentation slides - Foundation model comparison guide - XAI demonstration examples - Integration roadmap for Philippine agencies

Part 1: Geospatial Foundation Models (40 minutes)

What Are Foundation Models?

Foundation Models are large-scale AI models pre-trained on vast amounts of diverse data that can be adapted (fine-tuned) to many downstream tasks with minimal task-specific labeled data.

Analogy: Like a medical student who studies general anatomy before specializing in cardiology, a foundation model learns general patterns from millions of images before being fine-tuned for specific tasks like flood mapping or crop classification.

Key Characteristics: - Pre-trained on massive datasets (millions to billions of images) - General-purpose representations applicable to multiple tasks - Transfer learning enables fast adaptation with limited labels - Few-shot learning capability (learn from few examples)

TipWhy Foundation Models Matter for Philippines

Traditional Approach Problems: - Limited labeled Philippine satellite imagery - Models trained on US/European data don’t transfer well - Each new task requires thousands of labeled samples - Small agencies can’t afford massive labeling efforts

Foundation Model Solution: - Pre-trained on global satellite data (includes Philippine regions) - Fine-tune with just 100-500 labeled samples - One model → multiple applications - Democratizes advanced AI for resource-limited agencies

Major Geospatial Foundation Models

1. Prithvi (IBM-NASA, 2023)

  • Architecture: Temporal Vision Transformer (ViT)
  • Training Data: NASA’s Harmonized Landsat-Sentinel (HLS) data
    • 1 billion image patches
    • Multi-temporal sequences (time series)
    • 6 spectral bands
  • Pre-training Task: Masked Autoencoding
  • Applications: Flood mapping, wildfire detection, crop classification
  • Key Innovation: Handles temporal sequences, not just single images

2. Clay (Clay Foundation, 2024)

  • Architecture: Multi-modal Transformer
  • Training Data: Sentinel-1, Sentinel-2, Landsat, DEM
  • Pre-training Task: Contrastive learning + masked reconstruction
  • Applications: Land cover, change detection, disaster response
  • Key Innovation: Handles multiple sensor types simultaneously

3. SatMAE (Microsoft, 2022)

  • Architecture: Masked Autoencoder for satellite imagery
  • Training Data: Sentinel-2 imagery (10+ million images)
  • Pre-training Task: Mask 75% of patches, reconstruct
  • Applications: Scene classification, object detection
  • Key Innovation: Temporal masking captures seasonal patterns

4. DOFA (Geographic Foundation Model, 2024)

  • Architecture: Dynamic One-For-All (multi-task)
  • Training Data: Diverse EO datasets (optical + SAR)
  • Pre-training Task: Multi-task learning
  • Applications: Generalist model for any EO task
  • Key Innovation: Single model handles radar and optical

5. Google AlphaEarth (2024)

  • Architecture: Planetary-scale transformer (multimodal climate + EO)
  • Training Data: Global reanalysis, satellite observations, emissions inventories
  • Pre-training Task: Multi-task forecasting (temperature, precipitation, atmospheric constituents)
  • Applications: Climate risk analytics, long-range forecasting, emissions monitoring
  • Key Innovation: Couples climate dynamics with EO for actionable forecasts

How Foundation Models Work

flowchart TB
    subgraph PreTraining["Pre-Training Phase"]
        A[Massive Satellite Archive<br/>Millions of Images] --> B[Self-Supervised Learning<br/>No Labels Needed]
        B --> C[Foundation Model<br/>Learns General Features]
    end
    
    subgraph FineTuning["Fine-Tuning Phase"]
        C --> D[Your Small Labeled Dataset<br/>100-500 samples]
        D --> E[Task-Specific Model<br/>Flood/Drought/Crops]
    end
    
    subgraph Deployment["Deployment"]
        E --> F[Predict on New Data<br/>Operational Use]
    end
    
    style PreTraining fill:#e1f5ff
    style FineTuning fill:#fff4e1
    style Deployment fill:#e8f5e9

Comparison: Traditional vs. Foundation Model Approach

Aspect Traditional CNN Foundation Model
Training Data 10,000+ labeled samples 100-500 labeled samples
Training Time Days to weeks Hours
Compute Cost High (multi-GPU) Low (single GPU)
Performance Task-specific Multi-task capable
Generalization Limited Global knowledge

Philippine Use Cases

1. Rapid Disaster Response - Challenge: Flood maps within 6 hours of typhoon - Solution: Pre-trained Prithvi + 200 Philippine flood samples - Result: 15-minute flood extent generation

2. National Agricultural Monitoring - Challenge: Map crops for 12M hectares - Solution: Clay model (multi-sensor) + regional samples - Result: Nationwide coverage with transfer learning

3. Climate Resilience Planning - Challenge: Deliver seasonal flood/temperature forecasts with limited compute - Solution: Google AlphaEarth fine-tuned with PAGASA observations - Result: Actionable climate-risk layers feeding national adaptation plans

4. Mangrove Conservation - Challenge: Monitor 280K hectares on limited budget - Solution: SatMAE + 400 DENR samples - Result: Quarterly monitoring, minimal manual work

Part 2: Self-Supervised Learning (30 minutes)

What is Self-Supervised Learning?

Self-Supervised Learning (SSL) enables models to learn from unlabeled data by creating automatic tasks that require understanding data structure.

Key Idea: Instead of human labels, create “pretext tasks” that force the model to learn meaningful patterns.

Why SSL Matters for EO

The Labeling Problem: - Copernicus: 10 TB of satellite data per day - Only ~0.01% is labeled - Labeling cost: ₱50-200 per image - Domain experts are scarce

SSL Solution: - Learn from unlabeled data (99.99% available) - Reduces labeling needs by 10-100x - Transfers knowledge to labeled tasks

Common SSL Approaches

1. Masked Autoencoding (MAE)

Concept: Mask 60-75% of image patches, train model to reconstruct

Process: 1. Divide Sentinel-2 image into patches 2. Hide 75% randomly 3. Model predicts hidden patches 4. Learns spatial patterns, context, spectral signatures

flowchart LR
    A[Original Image] --> B[Mask 75%]
    B --> C[Encoder]
    C --> D[Decoder:<br/>Reconstruct]
    D --> E[Compare<br/>Calculate Loss]
    
    style A fill:#e8f5e9
    style B fill:#fff4e1
    style C fill:#e1f5ff

2. Contrastive Learning

Concept: Learn by comparing similar vs. dissimilar images

Process: 1. Take one image → create augmented versions 2. Train to recognize same image 3. Distinguish from other images 4. Learn invariant features

Applications: - Change detection - Multi-season crop mapping - Cross-sensor alignment

3. Temporal Self-Supervision

Concept: Predict temporal order or future states

Tasks: - Sort images by date - Predict next time step - Temporal interpolation

Value: Learns seasonal patterns crucial for agriculture

SSL Success: Philippine Mangrove Mapping

Traditional: 2,000 labels, 3 months, ₱300K SSL: Pre-train on 50K unlabeled + 200 labels, 3 weeks, ₱50K Result: Same accuracy, 80% cost reduction

Part 3: Explainable AI (XAI) (35 minutes)

Why Explainability Matters

The Black Box Problem: - Modern models are powerful but opaque - We know inputs/outputs, not reasoning - Critical for operational deployment

Philippine Scenarios Requiring XAI:

1. Disaster Response - NDRRMC prioritizes evacuations based on AI - Need: Explain flood risk predictions to LGUs

2. Agricultural Subsidies - DA allocates ₱10B in crop insurance - Need: Defend AI decisions, avoid bias

3. Environmental Enforcement - DENR detects illegal logging - Need: Legal evidence for prosecution

4. Scientific Credibility - PhilSA publishes forecasts - Need: Build trust with stakeholders

Major XAI Techniques

1. SHAP (SHapley Additive exPlanations)

What: Quantifies each feature’s contribution

Example: Drought LSTM

Predicted drought: 0.72 (high)

Contributions:
  NDVI (t-3):     -0.25  ← Low NDVI 3 months ago
  ONI (t-1):      +0.18  ← El Niño signal
  Rainfall (t-6): -0.12  ← Low rainfall 6 months ago
  
Final: 0.48 + 0.24 = 0.72

Value: Farmers understand why forecast was made

2. LIME (Local Interpretable Model-agnostic Explanations)

What: Creates simple local approximations

Process: 1. Take one prediction 2. Generate similar inputs 3. Fit simple linear model 4. Use coefficients as explanations

Use Case: Verify land cover classifications

3. Grad-CAM (Gradient-weighted Class Activation Mapping)

What: Visualizes CNN spatial attention

Output: Heatmap showing which image regions matter

Example: Building Detection - Red regions: Model focuses (rooftops, edges) - Blue regions: Model ignores (vegetation)

Use Case: Debug false positives

XAI Comparison

Technique Best For Output Limitation
SHAP Feature importance Numeric values Slow for large models
LIME Local explanations Simplified model Approximations
Grad-CAM Spatial attention Heatmap CNNs only

Part 4: Integration & Best Practices (15 minutes)

Decision Framework

When to use Foundation Models: - ✅ Limited labeled data (<1,000 samples) - ✅ Need rapid deployment - ✅ Task similar to pre-training data - ❌ Highly specialized Philippine-only task - ❌ Extreme data privacy concerns

When to use Self-Supervised Learning: - ✅ Abundant unlabeled data - ✅ Expensive labeling process - ✅ Want 5-10x label reduction - ❌ Already have large labeled set - ❌ Simple rule-based task

When to use Explainable AI: - ✅ High-stakes decisions (disaster, finance) - ✅ Need stakeholder trust - ✅ Regulatory requirements - ✅ Model debugging and improvement - ❌ Internal R&D only - ❌ Model already transparent (e.g., decision tree)

Implementation Roadmap for Philippine Agencies

Phase 1: Assessment (1-2 months) 1. Identify high-value use cases 2. Evaluate available labeled data 3. Select appropriate foundation model 4. Test on pilot region

Phase 2: Fine-Tuning (2-3 months) 1. Collect 200-500 Philippine samples 2. Fine-tune foundation model 3. Validate on held-out regions 4. Implement XAI for key predictions

Phase 3: Deployment (3-6 months) 1. Integrate with existing systems (DATOS, Space+) 2. Train staff on model interpretation 3. Establish monitoring and retraining schedule 4. Document for stakeholders

Phase 4: Scaling (6-12 months) 1. Expand to additional regions 2. Add new tasks using same foundation model 3. Build community of practice 4. Contribute Philippine data to global efforts

Resources and Next Steps

Foundation Model Repositories: - Prithvi: HuggingFace Model Hub - Clay: Clay Foundation - SatMAE: GitHub

XAI Tools: - SHAP: shap.readthedocs.io - LIME: github.com/marcotcr/lime - Captum (PyTorch): captum.ai

Philippine Platforms: - DOST-ASTI DATOS: asti.dost.gov.ph - PhilSA Space+ Dashboard - DIMER (Model Repository) - AIPI (AI Processing Interface)

Key Takeaways

ImportantSession 3 Summary

Foundation Models: - Pre-trained on billions of images globally - Fine-tune with 100-500 Philippine samples - Democratize advanced AI for small agencies - Reduce cost and time by 80-90%

Self-Supervised Learning: - Learn from unlabeled satellite archives - Reduce labeling needs by 10-100x - Critical for Philippine EO scaling - Powers foundation models

Explainable AI: - Essential for operational deployment - Build stakeholder trust - Debug model failures - Meet regulatory requirements - SHAP, LIME, Grad-CAM are key tools

Philippine Impact: - Overcome limited labeled data challenge - Rapid disaster response deployment - Cost-effective national monitoring - Transparent AI for government use

Looking Ahead to Session 4

NoteCourse Synthesis

Session 4 Topics: - Recap of all 4 days (Random Forest → Foundation Models) - Best practices for Philippine EO deployment - Building a Community of Practice - CoPhil Digital Space Campus - Your next steps in EO AI/ML

Reflection Questions for Session 4: 1. Which AI/ML technique fits your agency’s priority use case? 2. What are your biggest deployment barriers? 3. How can you contribute to the Philippine EO community?

Continue to Session 4 →

Additional Resources

Academic Papers: - Prithvi: Jakubik et al. (2023). “Foundation Models for Generalist Geospatial Artificial Intelligence” - SatMAE: Cong et al. (2022). “SatMAE: Pre-training Transformers for Temporal and Multi-Spectral Satellite Imagery” - XAI Survey: Lundberg & Lee (2017). “A Unified Approach to Interpreting Model Predictions”

Tutorials: - Fine-tuning Prithvi for Flood Mapping - SHAP for Time Series - Grad-CAM Implementation

Philippine Initiatives: - SkAI-Pinas Program - DIMER Model Repository - PhilSA-DOST Collaboration

This session is part of Day 4: Time Series Analysis, Emerging Trends, and Sustainable Learning - CoPhil 4-Day Advanced Training on AI/ML for Earth Observation, funded by the European Union under the Global Gateway initiative.