Session 1: Supervised Classification with Random Forest

Theory and Practice for Earth Observation

Stylianos Kotsopoulos

EU-Philippines CoPhil Programme

Session Overview

Part A: Theory (1.5 hours)

• Introduction to Supervised Classification
• Decision Trees Fundamentals
• Random Forest Ensemble Method
• Feature Importance
• Accuracy Assessment
• Google Earth Engine Platform

Part B: Hands-on Lab (1.5 hours)

• Sentinel-2 Data Acquisition
• Feature Engineering (Spectral Indices)
• Training Data Preparation
• Model Training & Optimization
• Classification & Validation
• Philippine NRM Applications

Learning Objectives:

• Understand supervised classification workflow for EO data
• Implement Random Forest using Google Earth Engine
• Perform accuracy assessment and interpret results
• Apply classification to Palawan land cover mapping

Part A: Theory

What is Supervised Classification?

• Goal: Assign labels to pixels/objects based on their characteristics
• “Supervised”: We provide labeled training examples to the algorithm
• Learning Process: Algorithm learns patterns from training data
• Application: Classify entire image based on learned patterns

Key Concept

Supervised classification requires labeled training data (ground truth) to learn the relationship between spectral signatures and land cover classes.

Supervised Classification Workflow

flowchart LR
    A[Satellite Imagery] --> B[Preprocessing]
    B --> C[Feature Extraction]
    C --> D[Training Data Collection]
    D --> E[Model Training]
    E --> F[Classification]
    F --> G[Accuracy Assessment]
    G --> H{Acceptable?}
    H -->|No| D
    H -->|Yes| I[Final Map]
    style E fill:#4A90E2
    style F fill:#4A90E2

Key Steps:

1. Preprocessing: Cloud masking, atmospheric correction
2. Feature Extraction: Spectral bands, indices (NDVI, NDWI)
3. Training Data: Collect representative samples for each class
4. Model Training: Train classifier on training data
5. Classification: Apply model to entire scene
6. Validation: Assess accuracy with independent test data

Common Land Cover Classes in Philippines

Natural Ecosystems:

• Primary Forest (dipterocarp)
• Secondary Forest
• Mangroves
• Grasslands
• Water Bodies (rivers, lakes, coastal)

Human-Modified:

• Agricultural Land (rice paddies, coconut)
• Urban/Built-up Areas
• Bare Soil
• Mining Areas
• Roads and Infrastructure

Philippine Context

Accurate land cover classification supports monitoring of Protected Areas, REDD+ programs, agricultural expansion, and disaster response (typhoons, floods).

Decision Trees

What is a Decision Tree?

A tree-like structure that makes decisions by asking a series of questions about features.

flowchart TD
    A[NDVI > 0.4?] -->|Yes| B[NIR > 3000?]
    A -->|No| C[NDWI > 0?]
    B -->|Yes| D[Forest]
    B -->|No| E[Agriculture]
    C -->|Yes| F[Water]
    C -->|No| G[Urban/Bare]
    style D fill:#2E7D32
    style E fill:#FBC02D
    style F fill:#1976D2
    style G fill:#757575

How it Works:

1. Start at root node
2. Test condition (e.g., NDVI > 0.4?)
3. Branch based on answer
4. Repeat until reaching leaf node
5. Leaf node = predicted class

Decision Tree Splitting

How does a tree decide where to split?

• Goal: Create pure nodes (all samples belong to one class)
• Metric: Information Gain or Gini Impurity
• Process: Test all possible splits, choose the best one
• Recursion: Repeat for each branch until stopping criteria

Gini Impurity Formula:

\[ Gini = 1 - \sum_{i=1}^{n} (p_i)^2 \]

Where \(p_i\) is the probability of class \(i\) in the node.

• Gini = 0: Pure node (all samples same class) ✓
• Gini = 0.5: Maximum impurity (50/50 split) ✗

Decision Tree Example: Spectral Splitting

• Split 1
• Split 2
• Split 3

Root Node: All 1000 samples

Test: NDVI > 0.4?

• Left branch (NDVI ≤ 0.4): 400 samples → Mostly Water, Urban, Bare
• Right branch (NDVI > 0.4): 600 samples → Mostly Forest, Agriculture

Information Gain: High ✓ (classes becoming more separated)

Right Branch: 600 samples with NDVI > 0.4

Test: NIR Reflectance > 3000?

• Left branch (NIR ≤ 3000): 250 samples → Agriculture (less canopy density)
• Right branch (NIR > 3000): 350 samples → Forest (dense canopy)

Information Gain: High ✓

Left Branch: 400 samples with NDVI ≤ 0.4

Test: NDWI > 0?

• Left branch (NDWI ≤ 0): 200 samples → Urban/Bare
• Right branch (NDWI > 0): 200 samples → Water

Information Gain: High ✓

Decision Tree Advantages & Limitations

Advantages:

• ✓ Easy to understand and visualize
• ✓ No data normalization needed
• ✓ Handles non-linear relationships
• ✓ Feature importance easily extracted
• ✓ Fast prediction

Limitations:

• ✗ Overfitting: Can memorize training data
• ✗ High variance: Small data changes → big tree changes
• ✗ Instability: Greedy algorithm (local optima)
• ✗ Bias: Favor features with many levels

The Solution

Random Forest addresses these limitations by combining many decision trees!

Random Forest

What is Random Forest?

An ensemble learning method that combines many decision trees to improve accuracy and reduce overfitting.

flowchart TD
    A[Training Data<br/>1000 samples] --> B1[Bootstrap<br/>Sample 1]
    A --> B2[Bootstrap<br/>Sample 2]
    A --> B3[Bootstrap<br/>Sample 3]
    A --> B4[...]
    A --> B5[Bootstrap<br/>Sample N]

    B1 --> T1[Tree 1]
    B2 --> T2[Tree 2]
    B3 --> T3[Tree 3]
    B4 --> T4[...]
    B5 --> T5[Tree N]

    T1 --> V[Majority Vote]
    T2 --> V
    T3 --> V
    T4 --> V
    T5 --> V

    V --> F[Final Prediction]

    style F fill:#4CAF50

Key Ideas:

1. Bootstrap Aggregating (Bagging)
   • Random sampling with replacement
   • Each tree sees different data
2. Random Feature Selection
   • Each split uses random subset of features
   • Reduces correlation between trees
3. Majority Voting
   • Classification: Most common class
   • Regression: Average prediction

Random Forest: The “Forest” Analogy

• One tree (Decision Tree): One expert’s opinion
  • Can be very confident but sometimes wrong
  • Might overfit to specific training examples
• Forest (Random Forest): Committee of experts
  • Each expert sees slightly different data
  • Each expert considers different features
  • Final decision: Majority vote
  • Wisdom of crowds: Group decision more reliable than individual

Intuition

If you ask 100 independent experts and 75 say “Forest”, you can be more confident than if only 1 expert says “Forest”.

Bootstrap Aggregating (Bagging)

Bootstrap Sampling:

• Original training set: 1000 samples
• Each tree gets: 1000 samples (with replacement)
• Some samples repeated, some never selected (~37% out-of-bag)

Original Data:

Sample IDs: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10

Tree 1 Bootstrap:

Sample IDs: 1, 3, 3, 5, 7, 7, 7, 9, 9, 10

Tree 2 Bootstrap:

Sample IDs: 2, 2, 4, 5, 5, 6, 8, 8, 9, 10

Why Bootstrap?

• Introduces diversity between trees
• Each tree specializes on different samples
• Reduces overfitting
• Enables Out-of-Bag (OOB) validation

Out-of-Bag Samples: - Samples not used by a tree (~37%) - Used for internal validation - No separate validation set needed

Random Feature Selection

At each split, only consider a random subset of features.

All Features (13 for Sentinel-2 + indices):

• B2 (Blue)
• B3 (Green)
• B4 (Red)
• B8 (NIR)
• B11 (SWIR1)
• B12 (SWIR2)
• NDVI
• NDWI
• NDBI
• EVI
• SAVI
• Texture features
• Elevation

Random Subset at Each Split:

Typical: \(\sqrt{n}\) features

For 13 features: \(\sqrt{13} \approx 4\) features

Tree 1, Split 1: {NDVI, B4, B11, Elevation}

Tree 1, Split 2: {B8, NDWI, B3, SAVI}

Tree 2, Split 1: {NDBI, B12, NDVI, B2}

. . .

Result: - Trees decorrelated - Prevents strong features from dominating - Better generalization

Random Forest Prediction: Majority Voting

Example Classification:

Classify a pixel with spectral signature: NDVI=0.65, NIR=4500, SWIR=2000

100 Trees Vote:

• Tree 1 → Forest 🌲
• Tree 2 → Forest 🌲
• Tree 3 → Agriculture 🌾
• Tree 4 → Forest 🌲
• Tree 5 → Forest 🌲
• …
• Tree 100 → Forest 🌲

Vote Count:

• Forest: 78 votes
• Agriculture: 22 votes

Final Prediction: Forest (78%)

Confidence: 78% confidence in prediction

Prediction Confidence

The proportion of votes can be interpreted as confidence. Higher consensus → more confident prediction.

Random Forest Hyperparameters

Key parameters to tune:

Parameter	Description	Typical Values	Impact
n_trees	Number of trees in forest	50-500	More trees → better performance (diminishing returns)
max_depth	Maximum depth of each tree	10-50 or None	Deeper → more complex, risk overfitting
min_samples_split	Min samples to split node	2-10	Higher → simpler trees, less overfitting
max_features	Features per split	\(\sqrt{n}\) or \(\log_2(n)\)	Balance between accuracy and diversity
bootstrap	Use bootstrap sampling	True	Almost always True for RF

Google Earth Engine Default

GEE’s ee.Classifier.smileRandomForest() defaults: - numberOfTrees: 100 - variablesPerSplit: \(\sqrt{n}\) (automatic) - minLeafPopulation: 1

Random Forest Advantages

1. High Accuracy
   • Often achieves excellent performance out-of-the-box
   • Handles complex non-linear relationships
2. Robust to Overfitting
   • Ensemble averaging reduces variance
   • Harder to overfit than single decision tree
3. Feature Importance
   • Quantifies which features matter most
   • Helps understand classification drivers
4. Handles Missing Data
   • Can work with incomplete feature sets
   • Robust to noisy data
5. No Normalization Needed
   • Works with features on different scales
   • Simplifies preprocessing
6. Efficient
   • Fast training (parallelizable)
   • Fast prediction

Feature Importance

Understanding Feature Importance

Question: Which spectral bands/indices contribute most to classification accuracy?

Feature Importance measures the contribution of each feature to the model’s predictions.

Calculation Methods:

1. Mean Decrease in Impurity (MDI)
   • How much each feature reduces impurity (Gini)
   • Averaged across all trees
   • Default in most implementations
2. Permutation Importance
   • Measure accuracy drop when feature is randomly shuffled
   • More reliable but slower

Interpretation:

• High importance: Feature strongly discriminates classes
• Low importance: Feature adds little information
• Zero importance: Feature not used by any tree

Example: Feature Importance for Palawan

Land Cover Classification (7 classes)

Top Features:

Rank	Feature	Importance	Use Case
1	NDVI	0.285	Forest vs. non-forest
2	NIR (B8)	0.192	Vegetation density
3	SWIR1 (B11)	0.156	Moisture content
4	NDWI	0.128	Water detection
5	Red (B4)	0.089	Vegetation health
6	NDBI	0.067	Urban areas
7	Elevation	0.045	Topographic context

Insights:

• NDVI dominant: Vegetation indices most important
• NIR crucial: Distinguishes vegetation types
• SWIR useful: Separates forest from agriculture
• NDWI essential: Water body identification
• Elevation helps: Mountains → forest, lowlands → agriculture

Actionable: - Focus on acquiring high-quality NIR and SWIR data - Ensure accurate NDVI calculation - Include DEM for improved accuracy

Feature Importance Visualization

Typical Output:

# Example feature importance from trained RF
Feature Importances:
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
NDVI        ████████████████████████████  0.285
NIR         ███████████████████           0.192
SWIR1       █████████████████             0.156
NDWI        ██████████████                0.128
Red         █████████                     0.089
NDBI        ███████                       0.067
Elevation   █████                         0.045
Blue        ███                           0.020
Green       ███                           0.018

Applications:

• Feature selection: Remove low-importance features
• Data collection priorities: Focus on important bands
• Model interpretation: Understand classification logic
• Domain validation: Does importance match EO theory?

Accuracy Assessment

Training/Test Data Splitting

Critical Decision: How to split data for training vs. validation?

Common Split Ratios:

Split	Training	Testing	Use Case
80/20	80%	20%	Standard (sufficient data)
70/30	70%	30%	More robust validation
60/40	60%	40%	Limited training data
50/50	50%	50%	Very small datasets

Google Earth Engine: Use .randomColumn() to assign splits

Splitting Strategies:

1. Random Split (most common)

# Add random column
data = data.randomColumn('random')

# Split 80/20
training = data.filter(ee.Filter.lt('random', 0.8))
testing = data.filter(ee.Filter.gte('random', 0.8))

2. Stratified Split (recommended) - Maintain class proportions in both sets - Important for imbalanced datasets - Ensures all classes in test set

3. Spatial Split - Training from one region, testing from another - Tests geographic transferability - More realistic for operational use

Random 80/20 is standard, but stratified ensures each class has representation in test set. Spatial split is most rigorous - tests if model works in new areas.

Why Accuracy Assessment?

• Quantify performance: How good is the classification?
• Compare models: Which classifier performs better?
• Identify weaknesses: Which classes are confused?
• Build confidence: Can we trust the map for decisions?
• Report to stakeholders: Scientific credibility

Golden Rule

ALWAYS use independent test data that was NOT used for training. Otherwise, you’re measuring memorization, not generalization.

Confusion Matrix

A table showing predicted classes vs. actual classes for test data.

Example: 5-class classification

	Predicted →	Forest	Agriculture	Water	Urban	Bare	Total	User’s Acc
Actual ↓
Forest		85	8	0	2	5	100	85%
Agriculture		12	73	0	5	10	100	73%
Water		0	1	95	2	2	100	95%
Urban		3	7	3	82	5	100	82%
Bare		5	11	2	9	78	105	74%
Total		105	100	100	100	100	505
Producer’s Acc		81%	73%	95%	82%	78%		OA: 82.6%

Accuracy Metrics Explained

• Overall Accuracy
• Producer’s Accuracy
• User’s Accuracy
• Kappa Coefficient

Definition: Percentage of correctly classified samples

\[ \text{Overall Accuracy} = \frac{\text{Correct Predictions}}{\text{Total Samples}} = \frac{85+73+95+82+78}{505} = \frac{413}{500} = 82.6\% \]

Interpretation: - Simple, intuitive metric - Limitation: Can be misleading with imbalanced classes

Example: 95% accuracy sounds great, but if 95% of pixels are forest, a “classify everything as forest” model achieves 95%!

Definition: Percentage of actual class correctly identified (per class)

Also called: Recall, Sensitivity

\[ \text{Producer's Accuracy}_{\text{Forest}} = \frac{85}{105} = 81\% \]

Interpretation: - “Of all actual forest pixels, how many did we correctly identify?” - Producer’s perspective: How complete is the map? - Low producer’s accuracy → omission error (missing true positives)

Example: 81% for Forest means we missed 19% of actual forest pixels.

Definition: Percentage of predicted class that is correct (per class)

Also called: Precision, Positive Predictive Value

\[ \text{User's Accuracy}_{\text{Forest}} = \frac{85}{100} = 85\% \]

Interpretation: - “Of all pixels we labeled as forest, how many are truly forest?” - User’s perspective: How reliable is the map? - Low user’s accuracy → commission error (false positives)

Example: 85% for Forest means 15% of “forest” pixels are actually other classes.

Definition: Measure of agreement beyond chance

\[ \kappa = \frac{p_o - p_e}{1 - p_e} \]

Where: - \(p_o\) = observed accuracy (overall accuracy) - \(p_e\) = expected accuracy by chance

Interpretation: - κ < 0.4: Poor agreement - κ = 0.4-0.6: Moderate agreement - κ = 0.6-0.8: Good agreement - κ > 0.8: Excellent agreement

Why use Kappa? Accounts for agreement by random chance, especially important for imbalanced datasets.

Common Confusion Patterns

Example: Forest vs. Agriculture confusion

	Predicted Forest	Predicted Agriculture
Actual Forest	85	8 ← Confusion
Actual Agriculture	12 ← Confusion	73

Why confusion occurs:

1. Spectral Similarity
   • Tree crops (coconut, fruit trees) look like forest
   • Young forest regeneration looks like agriculture
2. Mixed Pixels
   • Agroforestry systems
   • Forest edges with agriculture
3. Temporal Variability
   • Agriculture changes rapidly (planting, harvesting)
   • Single-date imagery may miss phenology
4. Class Definition Ambiguity
   • Where does “forest” end and “tree plantation” begin?

Best Practices: Training Data Collection

Practical Tips for High-Quality Training Samples:

Sample Size Guidelines:

Class	Min Samples	Recommended	Notes
Common (forest)	50	100-200	More coverage
Moderate (agriculture)	50	100-150	Capture variability
Rare (bare soil)	30	50-100	Get what you can

Sampling Strategies:

1. Stratified Random: - Distribute samples across study area - Avoid clustering in one region - Ensure all sub-types represented

2. Purposive Sampling: - Target known pure pixels - Use high-resolution imagery (Google Earth) - Field visits when possible

Quality Criteria:

✓ Pure Pixels - Homogeneous within polygon - Avoid edges and mixed areas - Use ≥3x3 pixel minimum areas

✓ Clear Definition - Unambiguous class membership - Document class definitions - Use consistent interpretation rules

✓ Temporal Match - Training data date matches imagery - Account for phenology (crops) - Update for multi-temporal analysis

Philippine-Specific Tips: - Use PhilSA Space+ Dashboard for recent imagery - Leverage NAMRIA land cover for reference - Consult LGU land use plans for urban areas - Use Google Street View for ground truth

Quality training data is more important than quantity. 100 high-quality samples beats 500 noisy samples. For Philippine applications, leverage existing government datasets as starting points.

Improving Classification Accuracy

• Better Training Data
• More Features
• Better Model
• Post-Processing

• More samples: 50-100 per class minimum
• Better quality: Pure pixels, clear boundaries
• Balanced: Equal samples per class
• Representative: Cover all variations within class
• Distributed: Spatial coverage across study area

• Multi-temporal data: Capture seasonal differences
• Texture features: Spatial patterns (GLCM)
• Topographic features: Elevation, slope, aspect
• Radar data: Sentinel-1 SAR (cloud-free)
• Ancillary data: Roads, protected areas

• Hyperparameter tuning: Optimize RF parameters
• More trees: 200-500 trees (diminishing returns after ~300)
• Cross-validation: Ensure generalization
• Ensemble methods: Combine RF with other classifiers

• Majority filter: Smooth noisy pixels
• Minimum mapping unit: Remove small isolated pixels
• Expert rules: Apply domain knowledge
• Visual inspection: Manual correction of obvious errors

Google Earth Engine

Why Google Earth Engine?

Challenges with Desktop GIS:

• ✗ Downloading large satellite data
• ✗ Storage requirements (TBs)
• ✗ Computational limitations
• ✗ Manual preprocessing
• ✗ Time-consuming workflows

Google Earth Engine Solutions:

• ✓ Petabyte-scale catalog (Landsat, Sentinel, MODIS…)
• ✓ Cloud computing (no downloads)
• ✓ Pre-processed data (atmospherically corrected)
• ✓ Scalable processing (parallel)
• ✓ Free for research & education

Perfect for This Course

GEE enables us to process years of Sentinel-2 data for entire provinces in minutes!

Google Earth Engine Architecture

flowchart TD
    A[User Code<br/>Python/JavaScript] --> B[GEE API]
    B --> C[GEE Cloud<br/>Processing]

    D[Satellite Data<br/>Catalog] --> C
    E[Landsat<br/>1972-present] --> D
    F[Sentinel-1/2<br/>2014-present] --> D
    G[MODIS<br/>2000-present] --> D
    H[Climate Data<br/>ERA5, etc.] --> D

    C --> I[Results]
    I --> J[Interactive Map]
    I --> K[Export to Drive]
    I --> L[Charts/Stats]

    style C fill:#4285F4
    style D fill:#34A853

Key Concepts:

• Server-side processing: Code runs on Google servers, not your laptop
• Lazy evaluation: Operations queued, executed only when needed (e.g., map display, export)
• Parallel processing: Automatically distributed across many machines

GEE Data Catalog Highlights

Relevant for Philippine EO:

Optical Imagery:

• Sentinel-2 MSI: 10m, 13 bands, 5-day revisit
• Landsat 8/9 OLI: 30m, 11 bands, 16-day revisit
• MODIS: 250-500m, daily, long time series

Radar:

• Sentinel-1 SAR: 10m, cloud-free, day/night

Terrain:

• SRTM DEM: 30m elevation
• ALOS World 3D: 30m (better for SE Asia)

Climate:

• ERA5: Hourly reanalysis (temp, precip)
• CHIRPS: Daily rainfall
• MODIS LST: Land surface temperature

Pre-processed Products:

• Hansen Global Forest Change: Annual tree cover loss
• ESA WorldCover: Global 10m land cover
• Global Surface Water: Water occurrence

GEE Code Editor vs. Python API

JavaScript (Code Editor)

• Pros:
  • Browser-based (no installation)
  • Interactive map interface
  • Built-in visualization
  • Great for exploration
• Cons:
  • Limited to GEE environment
  • Harder to integrate with other tools
  • Less powerful for data science

Use Case: Quick exploration, visualization

Python API

• Pros:
  • Integrate with NumPy, Pandas, scikit-learn
  • Jupyter notebooks
  • Reproducible workflows
  • Version control (Git)
  • Advanced analysis
• Cons:
  • Requires installation/setup
  • Slightly steeper learning curve

Use Case: Reproducible research, production workflows

Our Approach

We’ll use Python API with geemap library for best of both worlds: Python ecosystem + interactive maps!

GEE Random Forest Workflow

High-level workflow for today’s lab:

import ee
import geemap

# 1. Initialize GEE
ee.Initialize()

# 2. Load Sentinel-2 imagery
s2 = ee.ImageCollection('COPERNICUS/S2_SR') \
    .filterBounds(palawan_boundary) \
    .filterDate('2024-01-01', '2024-12-31') \
    .filter(ee.Filter.lt('CLOUDY_PIXEL_PERCENTAGE', 20))

# 3. Compute composite and indices
composite = s2.median()
ndvi = composite.normalizedDifference(['B8', 'B4']).rename('NDVI')
ndwi = composite.normalizedDifference(['B3', 'B8']).rename('NDWI')

# 4. Stack features
features = composite.select(['B2','B3','B4','B8','B11','B12']) \
    .addBands([ndvi, ndwi])

# 5. Sample training data
training = features.sampleRegions(collection=training_polygons,
                                   properties=['class'],
                                   scale=10)

# 6. Train Random Forest
classifier = ee.Classifier.smileRandomForest(numberOfTrees=100) \
    .train(features=training, classProperty='class', inputProperties=features.bandNames())

# 7. Classify image
classified = features.classify(classifier)

# 8. Visualize
Map = geemap.Map()
Map.addLayer(classified, vis_params, 'Land Cover')
Map

Philippine NRM Applications

Forest Monitoring (DENR, REDD+)

Challenges:

• 7,641 islands, 30 million hectares
• Cloud cover year-round
• Rapid deforestation in some areas
• Limited ground-based monitoring

RF Classification Helps:

• Annual forest cover maps
• Deforestation hotspot detection
• REDD+ MRV (Monitoring, Reporting, Verification)
• Protected area encroachment

Example: Palawan Biosphere Reserve

• Area: 1.1 million hectares
• Protection: UNESCO MAB, NIPAS
• Threats: Illegal logging, mining, agriculture

Workflow:

1. Annual Sentinel-2 composites (2016-2024)
2. RF classification (primary forest, secondary, non-forest)
3. Change detection (forest loss/gain)
4. Alert system for encroachment
5. Reports for PCSDS, DENR

Agricultural Monitoring (DA, PhilRice)

Rice Production Monitoring:

• Goal: Estimate planted area and yield
• Importance: Food security planning
• Traditional method: Field surveys (slow, expensive)

RF Approach:

• Multi-temporal Sentinel-2 (capture crop phenology)
• Training data from field surveys
• Classify: Rice, Other crops, Non-ag
• Area calculation per province/municipality
• Early warning for production shortfalls

Example: Central Luzon Rice Bowl

Classes: - Rice (wet season) - Rice (dry season) - Vegetables - Fallow/bare - Non-agricultural

Features: - NDVI time series (captures growth cycle) - LSWI (Land Surface Water Index) - EVI (Enhanced Vegetation Index)

Validation: - PhilRice field surveys - DA crop cut experiments

Urban Expansion Monitoring (NEDA, HLURB)

Metro Manila & Major Cities:

• Rapid urbanization: 3-5% annual growth
• Planning needs: Infrastructure, transport, housing
• Environmental concerns: Loss of green space, flooding

RF Classification:

• Urban/built-up
• Roads and infrastructure
• Vegetation (parks, trees)
• Bare soil (construction sites)
• Water bodies

Applications:

1. Urban growth tracking
   • Compare 2015 vs. 2024
   • Identify sprawl patterns
   • Predict future expansion
2. Green space monitoring
   • Urban vegetation loss
   • Park accessibility analysis
3. Flood risk
   • Impervious surface mapping
   • Drainage planning
4. Compliance
   • Illegal construction detection
   • Zoning violations

Water Resources (NWRB, LGUs)

Applications:

Surface Water Mapping:

• Rivers, lakes, reservoirs
• Seasonal variations
• Drought monitoring
• Flood extent mapping

RF Advantages: - NDWI as strong predictor - Multi-temporal captures seasonal changes - Can detect small water bodies

Watershed Management:

• Land cover within watersheds
• Forest cover (water regulation)
• Agriculture (erosion risk)
• Urban (runoff)

Example: Angat Dam Watershed - Critical for Metro Manila water supply - Monitor forest cover changes - Detect encroachment - Sediment risk assessment

Disaster Response (NDRRMC, PAGASA)

Post-Typhoon Damage Assessment:

Challenge: - Philippines: ~20 typhoons/year - Rapid assessment needed for relief - Cloud-free imagery rare after storms

RF Classification Approach:

1. Pre-event baseline: Land cover map
2. Post-event imagery: First clear Sentinel-2
3. Damage classes:
   • Intact forest/vegetation
   • Damaged vegetation
   • Exposed soil/landslides
   • Flooded areas
   • Building damage (requires very high res)

Example: Typhoon Odette (2021)

• Affected: Visayas, Mindanao
• Assessment needs:
  • Agricultural damage (coconut, rice)
  • Forest destruction
  • Coastal erosion
  • Flooded areas

RF Workflow: - Pre-typhoon: December 2021 composite - Post-typhoon: January 2022 composite - Classify: Intact, Damaged, Destroyed - Area statistics per municipality - Priority areas for relief

Recap: Session 1 Theory

What We Learned:

✓ Supervised classification workflow

✓ Decision trees: Intuitive but limited

✓ Random Forest: Ensemble of trees - Bootstrap sampling - Random feature selection - Majority voting

✓ Feature importance: Which bands matter?

✓ Accuracy assessment: - Confusion matrix - Overall, Producer’s, User’s accuracy - Kappa coefficient

✓ Google Earth Engine: Cloud-based EO

Key Takeaways:

1. Random Forest is powerful for EO classification
   • High accuracy
   • Handles non-linear relationships
   • Robust to overfitting
2. Training data quality is critical
   • Representative samples
   • Balanced classes
   • Sufficient quantity
3. Feature engineering improves results
   • Spectral indices (NDVI, NDWI)
   • Multi-temporal data
   • Auxiliary data (DEM)
4. Accuracy assessment builds confidence
   • Always use independent test data
   • Understand confusion patterns

Break

15-minute break before hands-on lab

🔸 Stretch

🔸 Coffee/water

🔸 Check your setup: - Google Earth Engine account - Python environment activated - Jupyter notebook ready

Coming up: Hands-on lab with Palawan land cover classification!

Part B: Hands-on Lab

Lab Overview

What We’ll Build: Palawan Land Cover Classification using Random Forest

Steps:

1. Setup and authentication
2. Load Sentinel-2 imagery
3. Create cloud-free composite
4. Calculate spectral indices
5. Prepare training data
6. Train Random Forest model
7. Generate classification map
8. Validate accuracy
9. Analyze results

Duration: ~1.5 hours

Study Area: Palawan Province

• Location: Western Philippines
• Area: ~14,649 km²
• Significance: UNESCO Biosphere Reserve
• Diversity: Forest, mangroves, agriculture, urban

Classes: 1. Forest 2. Agriculture 3. Water 4. Urban 5. Bare Soil

Spectral Indices for Classification

Key Features Beyond Raw Bands:

Vegetation Indices:

Index	Formula	Purpose
NDVI	(NIR - Red) / (NIR + Red)	Vegetation vigor
EVI	2.5 × (NIR - Red) / (NIR + 6×Red - 7.5×Blue + 1)	Enhanced sensitivity in high biomass

# Calculate NDVI
ndvi = image.normalizedDifference(['B8', 'B4']).rename('NDVI')

# Calculate EVI
evi = image.expression(
    '2.5 * ((NIR - RED) / (NIR + 6*RED - 7.5*BLUE + 1))',
    {'NIR': image.select('B8'),
     'RED': image.select('B4'),
     'BLUE': image.select('B2')
    }).rename('EVI')

Water & Built-up Indices:

Index	Formula	Purpose
NDWI	(Green - NIR) / (Green + NIR)	Water bodies
MNDWI	(Green - SWIR) / (Green + SWIR)	Water/wetlands (better separation)
NDBI	(SWIR - NIR) / (SWIR + NIR)	Built-up areas

# Calculate water indices
ndwi = image.normalizedDifference(['B3', 'B8']).rename('NDWI')
mndwi = image.normalizedDifference(['B3', 'B11']).rename('MNDWI')

# Calculate built-up index
ndbi = image.normalizedDifference(['B11', 'B8']).rename('NDBI')

Why MNDWI? Better separates water from built-up areas than NDWI (uses SWIR instead of NIR)

MNDWI is particularly useful in coastal areas like Palawan where you need to distinguish water bodies from urban areas. SWIR absorption by water creates stronger contrast.

Lab Instructions

Follow along in Jupyter notebook:

../notebooks/session1_hands_on_lab_student.ipynb

Student version: With TODO markers for exercises

Instructor version: Complete solutions

Tips for Success

• Read markdown cells carefully before running code
• Experiment with parameters
• Visualize intermediate results
• Ask questions when stuck!

Expected Outputs

By the end of the lab, you will have:

1. ✓ Interactive map of Palawan with Sentinel-2 composite
2. ✓ Calculated spectral indices (NDVI, NDWI, NDBI)
3. ✓ Trained Random Forest classifier (100 trees)
4. ✓ Land cover classification map
5. ✓ Confusion matrix and accuracy metrics
6. ✓ Feature importance ranking
7. ✓ Area statistics per land cover class
8. ✓ Exported classification to Google Drive

Accuracy Target: >80% overall accuracy

Session 1 Summary

Theory Concepts:

• Supervised classification workflow
• Decision trees → Random Forest
• Bootstrap aggregating
• Random feature selection
• Feature importance
• Accuracy assessment metrics
• Confusion matrix interpretation

Tools:

• Google Earth Engine
• Python API (geemap)
• Sentinel-2 imagery

Practical Skills:

• GEE authentication
• ImageCollection filtering
• Composite generation
• Spectral index calculation
• Training data preparation
• RF model training
• Classification execution
• Accuracy validation
• Map visualization

Philippine Context:

• Palawan land cover mapping
• DENR forest monitoring
• DA agricultural mapping
• NDRRMC disaster response

Next Session Preview

Session 2: Advanced Palawan Land Cover Lab

• Multi-temporal composites (dry/wet season)
• Advanced feature engineering (GLCM texture)
• Topographic features (DEM)
• 8-class detailed classification
• Hyperparameter tuning
• Change detection (2020 vs. 2024)
• Deforestation analysis
• Stakeholder reporting

Preparation:

• Complete Session 1 exercises
• Review confusion matrix analysis
• Think about classification improvements

Resources

Documentation:

• Google Earth Engine: https://developers.google.com/earth-engine
• geemap: https://geemap.org
• Sentinel-2: https://sentinel.esa.int/web/sentinel/missions/sentinel-2
• Random Forest paper: Breiman (2001) - Machine Learning 45:5-32

Philippine EO:

• PhilSA: https://philsa.gov.ph
• NAMRIA: https://namria.gov.ph
• DOST-ASTI PANDA: https://panda.stamina4space.upd.edu.ph

Course Materials:

• GitHub: [repository link]
• Datasets: [Google Drive link]

Thank You!

Questions?

Contact:

• Email: skotsopoulos@neuralio.ai
• Office Hours: [schedule]

Let’s move to the hands-on lab!

🚀 Open: session1_hands_on_lab.ipynb