Campus Land Cover Classification with Sentinel-2

This project performs binary land cover classification (Forest vs. Non-Forest) over a university campus area using Google Earth Engine and machine learning algorithms.

Overview

Using Sentinel-2 Surface Reflectance data, the project filters images to a strict 1% cloud cover threshold (QA60 masking) and evaluates three classifiers:

• Random Forest
• Support Vector Machine (SVM) (RBF Kernel)
• XGBoost (Gradient Boosted Trees)

Key Findings

• Internal validation (Campus bounds) demonstrates Random Forest as the strongest baseline (99.35%).
• External validation (Test Area unseen land) demonstrates that SVM and XGBoost tie for the highest generalization accuracy (91.67%).
• Winter imagery provides the most accurate spectral separability for the study area.
• 250 points per label were used for internal training/validation (70/30 split), and 240 points per label were evaluated externally.

Repository Structure

• fe/: Feature extraction and temporal/cloud sensitivity analysis notebooks.
• nb/:
  • rf/, svm/, xgb/: Classifier training, hyperparameter tuning, and spatial mapping notebooks.
  • cal_area/: Unseen test area validation notebooks for each classifier.
• assets/: Generated charts, heatmaps, and classified maps.

Setup Instructions

1. Install dependencies:

   pip install -r requirements.txt

2. Google Earth Engine Authentication: Authenticate your local Earth Engine environment:

   earthengine authenticate

   Ensure your .env file contains your Earth Engine Project ID:

   EE_PROJECT_ID=your-project-id

3. Run Notebooks: Open Jupyter notebooks to explore feature extraction, hyperparameter grid search, and final classification outputs.