Session 4: Synthesis, Q&A, and Pathway to Continued Learning

Bringing It All Together: From Basics to Operational AI for Philippine Earth Observation

Instructor

Stylianos Kotsopoulos

Date

November 17, 2025

Session 4: Course Synthesis & Your Next Steps

Consolidating 4 Days of Learning into Actionable Skills

From Python basics to foundation models - your journey to operational AI for Philippine Earth Observation

Session Overview

This final 2-hour session brings together everything you’ve learned across all four days. We’ll synthesize key AI/ML techniques, discuss best practices for operational deployment in Philippine agencies, introduce pathways for continued learning through the CoPhil Digital Space Campus, and foster a community of practice.

Learning Objectives

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

Synthesize all AI/ML techniques learned across 4 days
Map techniques to your agency’s priority use cases
Apply best practices for model deployment
Navigate the CoPhil Digital Space Campus
Engage with the Philippine EO AI community
Plan your next steps and projects

Presentation Slides

Session Details

Duration: 2 hours (120 minutes) | Format: Interactive Discussion + Q&A | Difficulty: Synthesis

Prerequisites: - Completion of Days 1-4 - Reflection on your agency’s use cases - Questions prepared for Q&A

Part 1: 4-Day Journey Recap (25 minutes)

The AI/ML Landscape We’ve Covered

flowchart TB
    subgraph Day1["Day 1: Foundations"]
        A1[Python for Geospatial]
        A2[Google Earth Engine]
        A3[Data Preprocessing]
    end
    
    subgraph Day2["Day 2: Machine Learning"]
        B1[Random Forest]
        B2[CNNs Introduction]
        B3[Land Cover Classification]
    end
    
    subgraph Day3["Day 3: Deep Learning"]
        C1[U-Net Segmentation]
        C2[Flood Mapping]
        C3[Object Detection]
    end
    
    subgraph Day4["Day 4: Advanced Topics"]
        D1[LSTMs]
        D2[Foundation Models]
        D3[XAI]
    end
    
    Day1 --> Day2
    Day2 --> Day3
    Day3 --> Day4
    
    style Day1 fill:#e8f5e9
    style Day2 fill:#e1f5ff
    style Day3 fill:#fff4e1
    style Day4 fill:#f3e5f5

Day-by-Day Synthesis

Day 1: Building the Foundation 🏗️

Core Skills: - Python for geospatial data (GeoPandas, Rasterio) - Google Earth Engine for satellite access - Sentinel-1 (SAR) and Sentinel-2 (Optical) data - Philippine EO ecosystem (PhilSA, NAMRIA, PAGASA)

Key Insight: Quality data preparation is 80% of the work

Day 2: Machine Learning Fundamentals 🌳

Techniques Mastered: - Random Forest for land cover classification - CNNs for automated feature extraction - Palawan land cover mapping case study

Key Insight: Random Forest still relevant - don’t always jump to deep learning

Philippine Applications: - NAMRIA land cover mapping - DENR forest classification - DA agricultural monitoring

Day 3: Deep Learning for Spatial Tasks �MAP️

Techniques Mastered: - U-Net for semantic segmentation - Central Luzon flood mapping - Pixel-wise classification - Object Detection (YOLO/SSD) - Metro Manila settlement detection

Key Insight: Deep learning shines with complex spatial patterns and sufficient data (1,000+ samples)

Philippine Applications: - NDRRMC rapid flood assessment - PhilSA urban growth monitoring - DENR mangrove mapping

Day 4: Time Series & Emerging Trends 🚀

Techniques Mastered: - LSTMs for temporal forecasting - Mindanao drought prediction - Foundation Models (Prithvi, Clay, SatMAE) - Fine-tune with 100-500 samples - Self-Supervised Learning - Learn from unlabeled data - Explainable AI (SHAP, LIME, Grad-CAM)

Key Insight: Foundation models + XAI democratize advanced AI for small agencies

Philippine Applications: - PAGASA seasonal forecasting - DA crop yield prediction - Multi-hazard monitoring

Technique Selection Matrix

Decision Tree: Which Technique to Use?

Question 1: What’s your prediction goal?

A. Single label per image → Image Classification - Small dataset (<1K): Transfer learning / foundation model - Large dataset (>5K): Train CNN or use foundation model

B. Pixel-wise labels → Semantic Segmentation - Best: U-Net or fine-tuned foundation model - Applications: Flood extent, land cover, vegetation

C. Bounding boxes → Object Detection - Use: YOLO, SSD, or DETR - Applications: Buildings, vehicles, ships

D. Time series → Sequence Modeling - Best: LSTM or Transformer - Applications: Drought, crop phenology, change detection

E. Tabular features → Traditional ML - Best: Random Forest or XGBoost - Advantages: Interpretable, fast

Question 2: How much labeled data?

<100: Foundation model fine-tuning
100-1,000: Transfer learning + augmentation
1,000-10,000: Smaller networks or traditional ML
>10,000: Full deep learning

Question 3: Need to explain predictions?

Yes (high-stakes): Use XAI or simpler models
No (research only): Any model

Part 2: Best Practices for Operational Deployment (30 minutes)

The Data-Centric AI Mindset

Traditional (Model-Centric): - Get any data → focus on model architecture → tune hyperparameters - Result: Marginal gains, high effort

Data-Centric Approach: - Systematically improve data quality → fix labeling errors → representative sampling - Result: Larger gains, sustainable performance

Andrew Ng’s 10-Minute Rule

Before spending 10 hours optimizing your model:

Spend 10 minutes examining training data
Look for:
- Labeling inconsistencies
- Missing edge cases
- Class imbalance
- Regional bias (e.g., all Luzon samples)
Fix data issues first
Then improve model

Case Study: Philippine flood model improved 78% → 91% accuracy by: - Fixing 50 mislabeled samples - Adding 100 Mindanao samples (was Luzon-only) - No model changes!

Validation Strategy Best Practices

1. Temporal Validation for Time Series

# WRONG: Random split (data leakage!)
train_test_split(X, y, test_size=0.2, random_state=42)

# RIGHT: Temporal split
train_end = '2020-12-31'
X_train = X[X.index <= train_end]
X_test = X[X.index > train_end]

2. Spatial Cross-Validation - Train on Luzon → Test on Visayas/Mindanao - Ensures model generalizes to unseen locations

3. Stratified Sampling - Proportional class representation - Diverse geographic locations - Seasonal variations included

Handling Class Imbalance

Problem: 95% non-flood, 5% flood pixels → Model predicts all non-flood → 95% accuracy but useless!

Solutions:

1. Weighted Loss

class_weights = {0: 1.0, 1: 19.0}  # non-flood : flood
model.compile(loss='weighted_cross_entropy', class_weight=class_weights)

2. Oversampling/Undersampling - Oversample minority (flood) - Undersample majority (non-flood) - Use SMOTE for synthetic samples

3. Better Metrics - Use F1-score, IoU, precision/recall - Focus on minority class performance

Deployment Checklist

Pre-Deployment Checklist ✅

Data Preparation: - [ ] Training data cleaned and verified - [ ] Validation strategy appropriate (temporal/spatial) - [ ] Test set represents deployment conditions - [ ] Class balance addressed

Model Development: - [ ] Baseline model established - [ ] Hyperparameter tuning completed - [ ] Multiple architectures compared - [ ] Model size appropriate for hardware

Validation: - [ ] Cross-validation performed - [ ] Test metrics meet requirements - [ ] Error analysis completed - [ ] Edge cases identified - [ ] XAI techniques applied

Operationalization: - [ ] Inference time acceptable - [ ] GPU/CPU requirements documented - [ ] API wrapper created - [ ] Monitoring dashboard set up - [ ] Retraining pipeline established

Stakeholder Communication: - [ ] Model limitations clearly stated - [ ] Uncertainty quantification provided - [ ] Explainability examples prepared - [ ] User training conducted - [ ] Feedback mechanism established

Model Monitoring & Maintenance

Models degrade over time due to: - Concept drift: Patterns change (climate change, urbanization) - Data drift: Input distribution shifts (new sensor, preprocessing) - Label drift: Definition changes (taxonomy update)

Solution: Continuous Monitoring

def monitor_model_performance():
    current_metrics = evaluate_on_latest_data()
    
    if current_metrics['accuracy'] < baseline - 0.05:
        alert_team("Model performance degraded!")
        trigger_retraining()
    
    log_metrics(current_metrics)

Retraining Schedule: - Disaster models: After each major event - Agricultural models: Quarterly (seasonal shifts) - Land cover models: Annually (gradual changes)

Part 3: CoPhil Digital Space Campus & Resources (20 minutes)

Your Continued Learning Platform

The CoPhil Digital Space Campus is your ongoing resource.

What’s Available:

1. Training Materials - All presentation slides (PDF/PPTX) - Jupyter notebooks (Colab-ready) - Datasets and examples - Code repositories - Video recordings (if available)

2. Self-Paced Learning - Beginner → Intermediate → Advanced tracks - Topic-specific modules - Hands-on labs with auto-grading - Capstone project templates

3. Resource Library - Curated research papers - Tool documentation - Case study repository - Philippine datasets catalog

Philippine EO AI Ecosystem

Key National Initiatives

1. SkAI-Pinas (Philippine Sky AI Program) - Lead: DOST-ASTI - Components: - DIMER: Model repository (share/download models) - AIPI: No-code/low-code ML platform

How to Engage: - Register at platform - Upload your models to DIMER - Try community models - Join monthly webinars

2. PhilSA Space+ Dashboard - National geospatial data portal - Sentinel-1, Sentinel-2 access - Pre-processed products (NDVI, LST) - API for programmatic access

3. NAMRIA Geoportal - Basemaps and boundaries - Hazard maps - Land cover datasets

4. PAGASA Climate Data Services - Historical weather records - Seasonal forecasts - El Niño/La Niña monitoring

External Resources

Foundation Models: - Prithvi: huggingface.co/ibm-nasa-geospatial - Clay: clay-foundation.github.io - SatMAE: github.com/sustainlab-group/SatMAE

Learning Platforms: - Google Earth Engine: earthengine.google.com - TensorFlow: tensorflow.org/tutorials - PyTorch: pytorch.org/tutorials - Fast.ai: fast.ai

Part 4: Building a Community of Practice (15 minutes)

Why Community Matters

Challenges of Solo Work: - Reinventing the wheel - Limited feedback - Difficulty staying current - Lack of domain best practices

Benefits of Community: - Knowledge sharing - Collaboration opportunities - Resource pooling - Collective advocacy

Philippine EO AI Community Activities

1. Monthly Virtual Meetups - Present your projects - Live coding sessions - Paper discussions - Tool demonstrations

2. Annual EO AI Summit - Showcase applications - International speakers - Networking - Hackathons

3. Collaborative Projects - Philippine Land Cover Map 2025 - Disaster AI Task Force - Agricultural Monitoring Network

4. Mentorship Program - Pair experienced with newcomers - Project guidance and code reviews

How to Contribute

Ways to Give Back

Share Your Work: - Upload models to DIMER - Write tutorials - Present at events - Publish case studies

Help Others: - Answer forum questions - Review code - Mentor newcomers - Organize study groups

Improve Resources: - Report issues - Suggest new topics - Contribute to open-source - Translate documentation

Advocate: - Promote EO AI in your agency - Encourage open data policies - Support capacity building - Bridge technical and policy communities

Part 5: Your Next Steps & Action Plan (20 minutes)

Immediate Actions (This Week)

1. Reflect on Your Learning - Which technique fits your agency’s needs? - What project could you start with? - What resources do you need?

2. Set Up Environment - Create Google Colab account - Bookmark CoPhil Campus - Register for SkAI-Pinas/DIMER - Join community forums

3. Start Small - Pick ONE simple project - Use existing notebooks as templates - Don’t aim for perfection - iterate!

Short-Term Goals (Next 3 Months)

Month 1: Proof of Concept - Select well-defined problem - Gather/access data - Implement baseline model - Document lessons learned

Month 2: Refinement - Address data quality issues - Try alternative approaches - Validate on multiple regions - Present to colleagues

Month 3: Mini-Deployment - Operationalize on pilot area - Create simple web app/report - Train end-users - Establish monitoring

Project Ideas by Agency

Starter Projects by Sector

Disaster Risk Reduction: - Real-time flood mapping (Sentinel-1) - Post-disaster damage assessment - Typhoon risk prediction - Landslide susceptibility updates

Agriculture: - Rice crop calendar monitoring - Drought early warning - Pest/disease outbreak detection - Crop type classification

Environment: - Forest cover change detection - Mangrove health monitoring - Mining activity surveillance - Protected area alerts

Urban Planning: - Informal settlement tracking - Urban heat island analysis - Traffic congestion monitoring - Building permit compliance

Water Resources: - Reservoir level monitoring - Irrigation assessment - Water quality proxies - Watershed mapping

Part 6: Open Q&A Session (10 minutes)

Common Questions

Q: “I don’t have GPU access. Can I still do deep learning?”

A: Yes! Strategies: - Google Colab (free GPU) - Smaller models and patches - Pre-trained models (less training) - University partnerships - Cloud credits (AWS, Google, Azure education)

Q: “How do I get more labeled data cheaply?”

A: Multiple approaches: - Crowdsourcing (students as annotators) - Semi-supervised learning - Active learning - Transfer learning - Synthetic data generation - Inter-agency data sharing

Q: “Our internet is slow. How to access large datasets?”

A: Workarounds: - CoPhil Mirror Site (local cache) - Process in cloud (GEE), download results - Pre-download during off-peak - Request hard drives from DOST/PhilSA - Focus on smaller AOIs initially

Q: “How to convince management to invest?”

A: Build business case: - Start with low-cost proof-of-concept - Quantify benefits (time/cost saved) - Show examples from other agencies - Emphasize national directives - Propose phased approach with milestones

Q: “What if my model doesn’t perform well?”

A: Systematic debugging: 1. Check data first (labels, balance, diversity) 2. Verify preprocessing (normalization, no leakage) 3. Simplify model (start with baseline) 4. Analyze errors (where/why fails?) 5. Try different approach 6. Ask community for help

Part 7: Course Conclusion (5 minutes)

What You’ve Achieved

In 4 days, you have: - ✅ Mastered Python for geospatial analysis - ✅ Learned to access Sentinel data (GEE) - ✅ Trained ML models (RF, CNN, U-Net, LSTM) - ✅ Applied to Philippine DRR, CCA, NRM challenges - ✅ Explored cutting-edge methods (Foundation Models, XAI) - ✅ Created operational-ready workflows - ✅ Joined a community of practice

You can now: - Propose and lead EO AI projects - Critically evaluate AI/ML solutions - Contribute to Philippine EO community - Continue learning independently

The Road Ahead

Remember: - Start small, iterate often - Focus on data quality - Explain your models (XAI builds trust) - Collaborate actively - Stay curious (AI evolves rapidly)

Final Thoughts

This training is the beginning of your journey as an EO AI practitioner.

The Philippines needs you: - To harness satellite data for disaster resilience - To support sustainable agriculture - To protect natural resources - To plan climate-adapted cities

You now have the skills to make an impact.

Go forth and build amazing things! 🚀🇵🇭

Course Feedback

Help Us Improve

Please complete: - 📝 Anonymous Feedback Form - 🎤 Brief exit interview (optional)

Tell us: - What worked well? - What needs improvement? - Topics to add/remove? - Pace and difficulty?

Certificates & Follow-Up

Digital Certificate: - Awarded upon 4-day completion - Download from CoPhil Campus - Includes competencies covered

Post-Training Support: - Email: skotsopoulos@neuralio.ai - Office hours: Bi-weekly

Alumni Network: - Join CoPhil Alumni groups - Annual reunion and showcase - Exclusive webinars

Final Reflection

Take 5 Minutes

Write down:

One technique you’ll apply first
One challenge you anticipate and how to address it
One person you’ll collaborate with or mentor
One goal for 3 months from now

Share with a peer if comfortable.

Closing Remarks

Congratulations! 🎉

You’ve completed the CoPhil 4-Day Advanced Training on AI/ML for Earth Observation.

What matters now is what you do next.

Take what you’ve learned, adapt it to your context, share with colleagues, and contribute to the community. Together, we’re building a more resilient, sustainable, and data-informed Philippines.

Mabuhay ang Philippine EO AI community! 🇵🇭🛰️🤖

Stay connected: - 🌐 CoPhil Digital Space Campus - 📧 skotsopoulos@neuralio.ai - 🤝 SkAI-Pinas Platform

Thank you for your dedication!

— The CoPhil Training Team

This session concludes 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.