flowchart TD
A[Session 1: RF Theory] --> B[Session 2: RF Practice]
B --> C[Session 3: CNN Theory]
C --> D[Session 4: CNN Practice]
A1[Decision Trees] --> A
A2[Ensemble Methods] --> A
A3[GEE Platform] --> A
B1[Advanced Features] --> B
B2[Multi-temporal] --> B
B3[Optimization] --> B
C1[Neural Networks] --> C
C2[Convolution] --> C
C3[Architectures] --> C
D1[TensorFlow/Keras] --> D
D2[Transfer Learning] --> D
D3[Evaluation] --> D
style A fill:#4A90E2
style B fill:#4A90E2
style C fill:#E85D75
style D fill:#E85D75
Day 2: Machine Learning for Earth Observation
From Random Forest to Convolutional Neural Networks
Day 2: Machine Learning for Earth Observation
From Random Forest to Convolutional Neural Networks
Master classical ML and deep learning techniques for satellite image analysis
Day 2 Overview
Day 2 bridges traditional machine learning and deep learning for Earth Observation applications. You’ll start with Random Forest classification using Google Earth Engine, then progress to Convolutional Neural Networks (CNNs) for advanced image analysis.
NoteWhat You’ll Learn
Morning (Classical ML): - Supervised classification with Random Forest - Land cover mapping with Sentinel-2 - Advanced feature engineering - Palawan case study with real-world applications
Afternoon (Deep Learning): - Neural network fundamentals - CNN architecture and operations - Building and training CNNs with TensorFlow/PyTorch - Image classification on EuroSAT dataset
Learning Objectives
By the end of Day 2, you will be able to:
- Implement supervised classification workflows using Random Forest in Google Earth Engine
- Engineer advanced features including GLCM texture, multi-temporal composites, and topographic data
- Optimize Random Forest models through hyperparameter tuning and cross-validation
- Perform comprehensive accuracy assessment using confusion matrices and per-class metrics
- Conduct change detection analysis and quantify deforestation rates
- Understand the transition from manual feature engineering to automatic feature learning
- Explain neural network fundamentals including forward/backward propagation and activation functions
- Comprehend CNN architecture components (convolution, pooling, dropout, dense layers)
- Build convolutional neural networks from scratch using TensorFlow/Keras
- Train deep learning models with advanced callbacks and GPU acceleration
- Apply transfer learning techniques using pre-trained models (ResNet, VGG, EfficientNet)
- Evaluate model performance using validation curves, confusion matrices, and error analysis
- Compare traditional machine learning vs deep learning approaches for EO applications
- Identify appropriate architectures for classification vs segmentation tasks
- Generate operational land cover maps and classification reports for stakeholder delivery
Today’s Schedule
| Time | Session | Topic | Materials |
|---|---|---|---|
| 09:00-12:00 | 1 | Random Forest Theory & Practice | Theory + Hands-on Lab |
| 13:00-15:00 | 2 | Palawan Land Cover Lab | Extended Hands-on Lab |
| 15:15-17:45 | 3 | Deep Learning & CNNs | Theory + Interactive |
| 18:00-20:30 | 4 | CNN Hands-on Lab | EuroSAT Classification |
Training Sessions
Session 1
Random Forest Theory & Practice
- Supervised classification workflow
- Decision trees and ensemble methods
- Google Earth Engine implementation
- Palawan land cover mapping
- Accuracy assessment
Session 2
Palawan Land Cover Lab
- Advanced feature engineering (GLCM, temporal)
- Multi-temporal composites
- 8-class detailed classification
- Hyperparameter tuning
- Change detection analysis
Session 3
Deep Learning & CNNs
- Neural network fundamentals
- CNN architecture explained
- Convolution and pooling operations
- Classic architectures (VGG, ResNet)
- CNNs for EO applications
Session 4
CNN Hands-on Lab
- EuroSAT image classification
- TensorFlow/Keras implementation
- PyTorch alternative track
- Transfer learning with ResNet
- Model evaluation & optimization
Presentations
Philippine Context: Palawan Case Study
Throughout Day 2, we focus on Palawan, a globally significant biodiversity hotspot:
Why Palawan?
- UNESCO Biosphere Reserve
- Critical habitat for endangered species
- Faces multiple environmental pressures
- Representative of broader Philippine challenges
Conservation Challenges:
- Deforestation from logging and agriculture
- Mining activities threatening ecosystems
- Coastal development impacting mangroves
- Climate change effects on biodiversity
- Balancing development and conservation
Monitoring Needs:
- Regular land cover mapping
- Deforestation detection and quantification
- Protected area encroachment monitoring
- Agricultural expansion tracking
- Reporting for REDD+ programs
Land Cover Classes:
- Primary forest (dipterocarp)
- Secondary forest
- Mangroves
- Agricultural land
- Grassland/scrubland
- Water bodies
- Urban/built-up
- Bare soil/mining
Stakeholders:
- DENR (forest management)
- PhilSA (EO monitoring)
- LGUs (local planning)
- NGOs (conservation)
- REDD+ programs
- Academic institutions
Technical Progression
Day 2 follows a deliberate pedagogical progression:
Morning → Afternoon Transition:
- Random Forest provides a strong foundation in supervised classification
- Understanding feature importance prepares for CNN’s automatic feature learning
- Confusion matrices and accuracy metrics carry over to deep learning evaluation
- GEE’s large-scale processing motivates need for efficient algorithms like CNNs
Prerequisites
From Day 1
You should have completed Day 1 or have equivalent knowledge:
Technical Setup
Recommended Review
- Sentinel-2 band combinations
- Spectral indices (NDVI, NDWI)
- Training vs. validation data
- Classification vs. regression
Hands-on Materials
All exercises run in Google Colaboratory (no local installation required):
📓 Jupyter Notebooks
Interactive coding environments with step-by-step instructions and exercises.
🗺️ Sample Datasets
Pre-prepared training data for Palawan and standardized benchmark datasets (EuroSAT).
💻 Code Templates
Reusable functions for common EO operations and analysis workflows.
📊 Visualizations
Interactive maps, charts, and diagrams to understand concepts and results.
Key Technologies
Classical ML: - Google Earth Engine (GEE) - geemap (Python API for GEE) - scikit-learn (Random Forest) - Sentinel-2 imagery
Deep Learning: - TensorFlow / Keras - PyTorch (alternative track) - EuroSAT dataset - GPU acceleration (Colab)
Data & Visualization: - NumPy, pandas - Matplotlib, seaborn - Rasterio, GeoPandas - Interactive mapping (folium, geemap)
Expected Outcomes
By the end of Day 2, you will have:
✅ Completed Projects: - Palawan land cover classification map (5-8 classes) - Change detection analysis (multi-temporal) - EuroSAT image classifier (10 classes)
✅ Technical Skills: - GEE Python API proficiency - Random Forest implementation and tuning - CNN architecture design - TensorFlow/Keras model building
✅ Practical Deliverables: - Working Jupyter notebooks - Trained models (RF and CNN) - Accuracy assessment reports - Exportable classification maps
Resources & Support
Need Help?
- Live instructor support during sessions
- Teaching assistants for breakout help
- Office hours (schedule TBA)
- Email: skotsopoulos@neuralio.ai
What’s Next?
ImportantReady to Begin?
Start with Session 1 to build your foundation in supervised classification:
Begin Session 1: Random Forest →
Check Your Readiness: - ✓ GEE account activated - ✓ Colab accessible - ✓ Day 1 concepts reviewed - ✓ Ready to code!
After Day 2:
Day 3 builds on these foundations with: - Advanced deep learning (U-Net for segmentation) - Object detection for EO - Time series analysis - Multi-modal data fusion
Day 2 is part of the CoPhil 4-Day Advanced Training on AI/ML for Earth Observation, funded by the European Union under the Global Gateway initiative and delivered in partnership with PhilSA and DOST.