Session 2: Advanced Palawan Land Cover Lab

Multi-temporal Classification and Change Detection

Instructor

CoPhil Advanced Training Program

Date

November 17, 2025

Session 2: Advanced Palawan Land Cover Lab

Multi-temporal Classification and Change Detection

Advanced feature engineering and real-world NRM applications

Session Overview

Duration: 2 hours | Type: Hands-on Lab | Difficulty: Intermediate

This session builds on Session 1 by implementing advanced classification techniques for a real-world Philippine conservation scenario: detailed land cover mapping of Palawan Biosphere Reserve.

Presentation Slides

ImportantPrerequisites
  • ✓ Complete Session 1 (Supervised Classification with Random Forest)
  • ✓ Google Earth Engine account authenticated
  • ✓ Python environment with geemap, ee, scikit-learn
  • ✓ Understanding of Random Forest basics

What You’ll Learn

After completing this session, you will be able to:

  1. Engineer Advanced Features
    • Calculate GLCM texture features (contrast, correlation, entropy)
    • Create multi-temporal composites
    • Integrate topographic data (DEM, slope, aspect)
    • Stack comprehensive feature sets
  2. Implement Multi-temporal Analysis
    • Process dry and wet season imagery
    • Calculate temporal indices and metrics
    • Detect seasonal patterns
  3. Optimize Classification Models
    • Perform hyperparameter tuning
    • Implement cross-validation strategies
    • Handle class imbalance
    • Apply post-processing techniques
  4. Conduct Change Detection
    • Compare multi-temporal classifications
    • Quantify deforestation rates
    • Identify change hotspots
    • Generate transition matrices
  5. Support NRM Decision-Making
    • Monitor protected areas
    • Detect encroachment
    • Generate stakeholder reports
    • Export results for GIS analysis

Lab Structure

Part A: Advanced Feature Engineering (30 minutes)

Learn to extract and combine multiple feature types for improved classification accuracy.

🌿 Texture Features (GLCM)

Calculate Gray-Level Co-occurrence Matrix features: - Contrast (local variation) - Correlation (pixel relationships) - Entropy (randomness) - Homogeneity (uniformity)

Useful for: Distinguishing forest types, urban texture

📅 Temporal Features

Create seasonal composites: - Dry season (Jan-May) - Wet season (Jun-Nov) - NDVI differences - Phenological signals

Useful for: Agriculture identification, seasonal wetlands

⛰️ Topographic Features

Extract from SRTM DEM: - Elevation - Slope - Aspect - Hillshade

Useful for: Forest/agriculture separation, land use patterns

🎯 Feature Stacking

Combine all features: - Spectral bands (6) - Indices (4) - Texture (4) - Temporal (4) - Topographic (3)

Total: ~20 features

Part B: Palawan Case Study (45 minutes)

Apply advanced classification to Palawan Biosphere Reserve using the 8-class scheme developed in Session 1.

Study Area: 11,655 km² UNESCO Biosphere Reserve

Classification Scheme:

Class Description Key Features
🌳 Primary Forest Dense dipterocarp, closed canopy High NDVI, low texture variation
🌲 Secondary Forest Regenerating, mixed canopy Moderate NDVI, medium texture
🌊 Mangroves Coastal, tidal influence High NDVI + high NDWI
🌾 Agricultural Rice, coconut plantations Seasonal NDVI patterns
🌿 Grassland Open areas, sparse vegetation Low-moderate NDVI
💧 Water Rivers, lakes, coastal Very low NIR, high NDWI
🏘️ Urban Settlements, infrastructure High NDBI, low NDVI
⛏️ Bare Soil Mining, cleared land Bright reflectance, low NDVI

Part C: Model Optimization (30 minutes)

Fine-tune your Random Forest classifier for maximum accuracy.

Optimization Techniques:

  1. Hyperparameter Tuning
    • Number of trees (50, 100, 200, 500)
    • Variables per split (sqrt, log2, all)
    • Minimum samples per leaf (1, 2, 5)
  2. Cross-Validation
    • K-fold validation (k=5)
    • Stratified sampling
    • Out-of-bag error estimation
  3. Class Balancing
    • Handle imbalanced classes
    • Weighted training samples
    • SMOTE (if needed)
  4. Post-Processing
    • Majority filtering (reduce salt-and-pepper noise)
    • Minimum mapping unit enforcement
    • Edge smoothing

Part D: NRM Applications (15 minutes)

Apply your classification to real conservation challenges in Palawan.

🚨 Deforestation Detection

Compare 2020 vs 2024: - Forest loss hotspots - Conversion patterns - Quantify area changes - Generate alerts

🌾 Agricultural Expansion

Track land conversion: - Forest → Agriculture - Grassland → Agriculture - Expansion rates - Proximity to roads

🛡️ Protected Area Monitoring

Assess threats: - Boundary encroachment - Internal degradation - Buffer zone changes - Compliance tracking

📊 Stakeholder Reports

Generate outputs: - Area statistics by class - Change matrices - Maps (GeoTIFF, PNG) - CSV summary tables

Key Concepts

GLCM Texture Analysis

What is GLCM?
Gray-Level Co-occurrence Matrix measures spatial relationships between pixel pairs, capturing image texture.

Why use it? - Distinguishes primary vs secondary forest (canopy structure) - Separates urban from bare soil (heterogeneity) - Identifies mangrove stands (unique texture)

GEE Implementation:

# Add NIR texture features
texture = image.select('B8').glcmTexture(size=3)
contrast = texture.select('B8_contrast')
entropy = texture.select('B8_ent')

Multi-temporal Composites

Philippine Seasons: - Dry (Dec-May): Best for forest mapping, less cloud cover - Wet (Jun-Nov): Shows maximum vegetation, agricultural phenology

Temporal Indices: - NDVI Difference: Wet NDVI - Dry NDVI
Positive: Seasonal crops (rice)
Near zero: Evergreen forest
Negative: Dry season crops

Benefits: - Reduce cloud impacts (median compositing) - Capture phenological cycles - Improve agricultural separation - Detect irrigated vs rainfed

Hyperparameter Tuning

Key Random Forest Parameters:

Parameter Effect Recommended Range
numberOfTrees More trees = better but slower 100-500
variablesPerSplit Features per split sqrt(n) for classification
minLeafPopulation Minimum samples in leaf 1-5
bagFraction Training sample fraction 0.5-0.7

Optimization Strategy: 1. Start with defaults 2. Grid search on key parameters 3. Use out-of-bag error for evaluation 4. Validate on independent test set

Palawan Conservation Context

Why Palawan Matters

Biodiversity Hotspot: - 252 bird species (15 endemic) - 95 mammal species - Last Philippine frontier forest - Critically endangered species habitat

Threats: - Mining (nickel, chromite) - Agricultural expansion - Infrastructure development - Illegal logging - Tourism pressure

Conservation Status: - UNESCO Biosphere Reserve (1990) - Multiple protected areas - Strategic Environmental Plan (SEP) framework - National government priority

NRM Applications

DENR Monitoring: - Annual forest cover updates - REDD+ MRV requirements - Protected area assessments - Permit compliance checking

Local Government: - Land use planning - Infrastructure siting - Agricultural zoning - Disaster risk assessment

NGO Conservation: - Deforestation alerts - Community monitoring - Baseline assessments - Impact evaluation

Hands-On Notebook

Access the Lab

Tip📓 Jupyter Notebook

The complete hands-on lab is available as an interactive Jupyter notebook:

Student Version (with exercises):
session2_extended_lab_STUDENT.ipynb

Instructor Version (with solutions):
session2_extended_lab_INSTRUCTOR.ipynb

Google Colab:
Open In Colab

Code Templates

Reusable Python functions for advanced features:

  1. GLCM Texture: glcm_template.py
  2. Temporal Composites: temporal_composite_template.py
  3. Change Detection: change_detection_template.py

Expected Outcomes

Classification Performance

Target Accuracy Metrics: - Overall Accuracy: >85% - Kappa Coefficient: >0.80 - Per-class accuracy: >80% for most classes

Common Confusion: - Primary ↔︎ Secondary Forest (canopy density gradient) - Mangroves ↔︎ Agriculture (wet season similarity) - Urban ↔︎ Bare Soil (bright surfaces)

Deliverables

By the end of this session, you will produce:

✅ High-resolution land cover map (10m)
✅ Accuracy assessment report
✅ Feature importance analysis
✅ 2020-2024 change detection map
✅ Area statistics by class
✅ Deforestation hotspot map
✅ Exported GeoTIFF for GIS

Troubleshooting

Common Issues

“Computation timed out” - Reduce study area size - Use smaller GLCM window (3x3 instead of 5x5) - Process in tiles

“Memory limit exceeded” - Export intermediate results - Use .aside() sparingly - Reduce feature count

“Low classification accuracy” - Check training data quality - Add more training samples - Try different feature combinations - Adjust class definitions

“GEE authentication failed”

import ee
ee.Authenticate()
ee.Initialize()

Getting Help

Additional Resources

Documentation

Datasets

Scientific Papers

  • Karra et al. (2021). Global LULC with Sentinel-2 and deep learning. IGARSS
  • Phiri et al. (2020). Sentinel-2 LULC classification: A review. Remote Sensing
  • Belgiu & Drăguţ (2016). Random Forest in remote sensing. ISPRS

Assessment

Formative Assessment

  • ✓ Complete all TODO exercises in notebook
  • ✓ Achieve >80% classification accuracy
  • ✓ Generate all required outputs
  • ✓ Answer concept check questions

Summative Assessment

  • Classification map quality (40%)
  • Accuracy metrics achieved (30%)
  • Change detection analysis (20%)
  • Written interpretation (10%)

Next Steps

NoteAfter Session 2

You’re now ready for deep learning approaches!

Session 3 introduces Convolutional Neural Networks (CNNs) for Earth observation, building on your classification experience.

Continue to Session 3 →

Extended Projects

Want to go further? Try these:

  1. Expand Study Area: Apply to entire Palawan or other Philippine regions
  2. Add Classes: Separate coconut vs rice, primary forest sub-types
  3. Time Series: Analyze annual trends (2017-2024)
  4. Integration: Combine with field data or local knowledge
  5. Automation: Build monitoring pipeline with regular updates