Land Use / Land Cover (LULC) Classification using Google Earth Engine

This project focuses on Land Use and Land Cover (LULC) classification of Uttarakhand/Uttarkashi using Google Earth Engine (GEE), Sentinel-2 satellite imagery, spectral indices, and Random Forest classification techniques.

The study integrates open-source global LULC datasets such as ESA WorldCover and Dynamic World for comparative analysis and validation.

---

Objectives

• To perform LULC classification using Sentinel-2 imagery.
• To derive spectral indices such as:
  • NDVI (Normalized Difference Vegetation Index)
  • NDWI (Normalized Difference Water Index)
• To improve class separability and classification accuracy using spectral indices.
• To utilize ESA WorldCover and Dynamic World datasets for baseline comparison and validation.
• To assess classification performance using:
  • Confusion Matrix
  • Overall Accuracy
  • Kappa Coefficient

---

Study Area

• Uttarakhand, India
• Uttarkashi District (for detailed classification)

---

Datasets Used

Sentinel-2 Surface Reflectance

Dataset:

COPERNICUS/S2_SR_HARMONIZED

Used Bands:

• B2 → Blue
• B3 → Green
• B4 → Red
• B8 → Near Infrared (NIR)

---

ESA WorldCover

Dataset:

ESA/WorldCover/v200

---

Dynamic World

Dataset:

GOOGLE/DYNAMICWORLD/V1

---

Spectral Indices

NDVI

Used for vegetation detection.

[ NDVI = \frac{B8 - B4}{B8 + B4} ]

---

NDWI

Used for water body extraction.

[ NDWI = \frac{B3 - B8}{B3 + B8} ]

---

Methodology

1. Load study area boundary.
2. Import Sentinel-2 imagery.
3. Apply cloud filtering and image compositing.
4. Generate NDVI and NDWI.
5. Create training polygons manually.
6. Sample training data.
7. Split data into:
   • Training set
   • Validation set
8. Train Random Forest classifier.
9. Perform LULC classification.
10. Assess accuracy using confusion matrix.

---

Classification Classes

Class	Label
0	Water
1	Vegetation
2	Built-up
3	Bare Land

---

Accuracy Assessment

The model performance was evaluated using:

• Overall Accuracy
• Kappa Coefficient
• Producer Accuracy
• User Accuracy

Example Results:

• Overall Accuracy ≈ 95%
• Kappa ≈ 0.91

---

Project Structure

LandUse_LandCover/
│
├── Results/
│   ├── Accuracy_assm.png
│   ├── adm_bound.jpeg
│   ├── Dynamic_world.png
│   ├── ESA_world_cover.png
│   ├── NDVI.png
│   ├── NDWI.png
│   └── train_samples.png
│
├── accuracy_ass.js
├── adm_boundary.js
├── esa_world_cover.js
├── NDVI_NDWI.js
├── random_forest.js
├── sat_data.js
└── train_valid.js

---

Software & Tools

• Google Earth Engine (GEE)
• JavaScript API
• Visual Studio Code
• Git & GitHub

---

Results

The classification successfully identified:

• Vegetation cover
• Water bodies
• Built-up areas
• Bare land regions

Spectral indices significantly improved class separability and classification accuracy in mountainous terrain.

---

Future Improvements

• Addition of DEM and slope datasets
• Multi-temporal analysis
• Deep learning-based classification
• Higher-resolution validation datasets

---

Author

Kajal Rajput
NSUT Geoinformatics Engineering
Batch 2023–2027

---

References

• Sentinel-2: European Space Agency (ESA)
• Google Earth Engine Documentation
• ESA WorldCover Dataset
• Dynamic World Dataset by Google