Session 1: Supervised Classification with Random Forest

Theory and Practice for Earth Observation

Instructor

CoPhil Advanced Training Program

Date

November 17, 2025

Session 1: Supervised Classification with Random Forest

Theory and Practice for Earth Observation

Learn to classify satellite imagery using Random Forest and Google Earth Engine

Session Overview

This 3-hour session introduces supervised classification for Earth Observation, focusing on the Random Forest algorithm and its implementation in Google Earth Engine.

Presentation

TipInteractive Reveal.js Slides

Open Slides →

NoteLearning Objectives

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

  1. Understand supervised classification workflow for EO data
  2. Explain how decision trees and Random Forest work
  3. Implement Random Forest classification using Google Earth Engine
  4. Perform accuracy assessment and interpret results
  5. Apply classification to Palawan land cover mapping

Presentation Slides

Part A: Theory (1.5 hours)

Topics Covered

Supervised Classification Basics

  • Classification workflow
  • Training data requirements
  • Feature extraction
  • Model validation

Decision Trees

  • Recursive partitioning
  • Splitting criteria (Gini, entropy)
  • Overfitting prevention
  • Tree visualization

Random Forest Ensemble

  • Bootstrap aggregation (bagging)
  • Random feature selection
  • Voting mechanism
  • Ensemble advantages

Feature Importance

  • Variable selection
  • Spectral band importance
  • Index contribution
  • Interpretation strategies

Accuracy Assessment

  • Confusion matrices
  • Producer’s accuracy (recall)
  • User’s accuracy (precision)
  • Kappa coefficient

Google Earth Engine

  • Platform capabilities
  • Cloud computing advantages
  • GEE Python API
  • geemap library

Theory Materials

📊 Interactive Theory Notebook

Step-by-step exploration of Random Forest concepts with visualizations and exercises.

  • Decision tree interactive demos
  • Random Forest voting visualization
  • Feature importance analysis
  • Confusion matrix interpretation
  • Concept check quiz

Open Theory Notebook

⏱️ Estimated Time: 70 minutes

Part B: Hands-on Lab (1.5 hours)

Practical Workflow

This hands-on lab guides you through a complete land cover classification workflow for Palawan using Sentinel-2 imagery and Random Forest.

Lab Sections

1. Setup (10 minutes) - Import libraries (geemap, ee, scikit-learn) - Authenticate Google Earth Engine - Define Palawan study area - Set date range for imagery

2. Data Acquisition (20 minutes) - Load Sentinel-2 ImageCollection - Apply cloud masking (QA60 band) - Filter by date and cloud cover - Compute spectral indices: - NDVI (vegetation) - NDWI (water) - NDBI (built-up) - EVI (enhanced vegetation) - Create median composite - Visualize RGB and false color

3. Training Data (20 minutes) - Load training polygons for land cover classes: - Forest - Agriculture - Water - Urban - Bare soil - Visualize training areas on map - Sample spectral values - Prepare feature matrix - Inspect training statistics

4. Model Training (20 minutes) - Configure Random Forest classifier - Set hyperparameters: - Number of trees (e.g., 100) - Variables per split - Min leaf population - Train model on Sentinel-2 features - Inspect feature importance

5. Classification (15 minutes) - Apply trained model to study area - Generate land cover map - Create visualization palette - Display results on interactive map - Calculate area statistics by class

6. Validation (25 minutes) - Split data into train/test (80/20) - Apply model to validation data - Calculate accuracy metrics: - Overall accuracy - Producer’s accuracy - User’s accuracy - Kappa coefficient - Generate confusion matrix - Analyze misclassification patterns

7. Exercises (20 minutes) - Modify number of trees - Add/remove spectral indices - Classify different Philippine region - Export results to Google Drive

Lab Materials

💻 Hands-on Lab Notebook

Complete implementation of Random Forest classification workflow with Google Earth Engine.

Open Lab Notebook

⏱️ Estimated Time: 120 minutes

📋 Prerequisites: - Google Earth Engine account (sign up) - Basic Python knowledge - Understanding of remote sensing concepts

Philippine Context

This session uses Palawan as the study area, a biodiversity hotspot and UNESCO Biosphere Reserve facing challenges from:

  • Deforestation and illegal logging
  • Agricultural expansion
  • Mining activities
  • Coastal development
  • Climate change impacts

Applications for Philippine Agencies:

  • DENR: Forest monitoring and protected area management
  • DA: Agricultural land mapping and crop monitoring
  • NDRRMC: Land cover baseline for disaster risk assessment
  • PhilSA: Operational land cover products
  • LGUs: Municipal planning and zoning

Key Concepts

What is Supervised Classification?

Supervised classification assigns labels to pixels based on their spectral characteristics, using labeled training data to learn the relationship between spectral signatures and land cover classes.

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

Why Random Forest for EO?

TipRandom Forest Advantages

Strengths: - Handles high-dimensional data (many spectral bands/indices) - Resistant to overfitting through ensemble averaging - Provides feature importance rankings - Fast training and prediction - No assumptions about data distribution - Works well with limited training samples

Ideal for: - Land cover classification - Crop type mapping - Forest/non-forest discrimination - Urban area detection

Common Land Cover Classes

Natural Ecosystems: - Primary forest (dipterocarp) - Secondary forest - Mangroves - Grasslands - Water bodies

Human-Modified: - Agricultural land (rice, coconut) - Urban/built-up areas - Bare soil - Mining areas - Roads and infrastructure

Technical Details

Google Earth Engine

Google Earth Engine provides:

  • Petabyte-scale satellite imagery archive
  • Cloud computing infrastructure
  • Pre-processed analysis-ready data
  • JavaScript and Python APIs
  • Free access for research and education

Key GEE Concepts: - ImageCollection: Time series of satellite images - Image: Single satellite scene or composite - Feature/FeatureCollection: Vector data (training polygons) - Reducer: Aggregation operations - Classifier: Machine learning algorithms

Random Forest Hyperparameters

Parameter Description Typical Range Effect
numberOfTrees Number of trees in ensemble 50-500 More trees → better performance, longer training
variablesPerSplit Features to consider per split √n to n Lower → more randomness, less correlation
minLeafPopulation Min samples in leaf node 1-10 Higher → simpler trees, less overfitting
bagFraction Fraction for bootstrapping 0.5-1.0 Lower → more randomness
maxNodes Max nodes per tree null (unlimited) Limits tree complexity

Spectral Indices for Classification

Vegetation Indices: - NDVI: (NIR - Red) / (NIR + Red) → vegetation vigor - EVI: 2.5 × (NIR - Red) / (NIR + 6×Red - 7.5×Blue + 1) → improved sensitivity

Water Indices: - NDWI: (Green - NIR) / (Green + NIR) → water bodies - MNDWI: (Green - SWIR) / (Green + SWIR) → water/wetlands

Built-up Indices: - NDBI: (SWIR - NIR) / (SWIR + NIR) → urban areas

Resources

Documentation

Philippine EO Resources

Course Materials

Theory Notebook Lab Notebook Setup Guide FAQ

Next Steps

ImportantPreparation for Session 2

Session 2 builds on these fundamentals with advanced Palawan land cover classification:

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

To Prepare: - Complete all Session 1 exercises - Review confusion matrix analysis - Think about potential classification improvements - Consider additional features that might help

Preview Session 2 →

Questions & Support

During Session: - Ask questions in chat or unmute - Use breakout rooms for small group help - Teaching assistants available

After Session: - Review FAQ - Email instructors: skotsopoulos@neuralio.ai - Office hours: [schedule]

This session 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.