🛰️ LULC Classification — East Sikkim, India

Summer Training Programme on Remote Sensing and GIS – 2025
India Space Academy | Department of Space Education

Prepared by: Satwik Shreshth  |  MCA 2nd Year, Sikkim University (Central University)

---

📌 Overview

This project generates a 10-metre resolution Land Use Land Cover (LULC) map of East Sikkim district, Sikkim, India using supervised ensemble machine learning classifiers trained on multi-temporal Sentinel-2 Surface Reflectance imagery. All satellite processing was performed on the Google Earth Engine (GEE) cloud computing platform; accuracy evaluation and visualisation were done in Python (scikit-learn) and QGIS.

Two classifiers are compared — Random Forest (RF) and Gradient Tree Boosting (GTB) — across five thematic land cover classes. RF was selected for the final map export.

---

🗺️ LULC Map

10-m resolution LULC map of East Sikkim — Random Forest classifier, visualised in QGIS.

Symbol	Class
🔵 Blue	Water
🟢 Dark Green	Forest
⬜ White	Highland / Snow Cover
🔴 Red	Built-up Area
Greyish Pink	Barren Land

---

📂 Repository Structure

LULC_Classification/
│
├── 📁 EastSikkimSHP/                    # District boundary shapefile (FAO GAUL 2015)
│   ├── AOI_EastSikkim.shp
│   ├── AOI_EastSikkim.shx
│   ├── AOI_EastSikkim.dbf
│   └── AOI_EastSikkim.prj
│
├── 📁 TrainingPoint Gangtok/            # Manual ground-truth training points (GEE asset)
│
├── 📁 eval_plots_LULC_EastSikkim/       # All evaluation plots
│   ├── confusion_matrix_RF.png
│   ├── confusion_matrix_GTB.png
│   ├── roc_curves_RF.png
│   ├── roc_curves_GTB.png
│   ├── f1_comparison.png
│   ├── omission_commission.png
│   └── overall_accuracy_kappa.png
│
├── 📓 LULC.ipynb                        # Jupyter Notebook — Preprocessing, Feature Engineering, Training 	accuracy assessment & plots
├── 🖼️  LULC_East_Sikkim.png             # Final LULC map (PNG, QGIS export)
├── 🗂️  LULC_East_Sikkim_Clipped.tif     # Full-resolution clipped GeoTIFF
├── 🗂️  LULC_EastSikkim_RF.tif           # Random Forest classified raster (GeoTIFF)
├── 📦 eval_plots_LULC_EastSikkim.zip    # Compressed evaluation plots archive
├── 📄 report_data.json                  # Accuracy metrics and evaluation data
├── 📑 Report.pdf                        # Full project report
└── 📝 README.md                         # This file

---

🔬 Methodology

Study Area

Parameter	Value
District	East Sikkim, Sikkim, India
Area	~1,521.8 km²
Latitude	27.14°N – 27.42°N
Longitude	88.44°E – 88.92°E
Elevation Range	~300 m to >4,000 m
Coordinate System	WGS84 / EPSG:4326

Satellite Data & Pre-processing

• Source: COPERNICUS/S2_SR_HARMONIZED on Google Earth Engine
• Period: January 2020 – January 2024 | Cloud filter: <80%
• Cloud Masking: QA60 bitmask (bits 10 & 11); snow pixels preserved where NDSI > 0.40
• Composite: Pixel-wise median clipped to AOI

Feature Stack (15 Features)

Category	Features
Spectral Bands	B2, B3, B4, B8, B11, B12
Spectral Indices	NDVI, NDWI, NDSI
Phenological	NDVIwin (Dec–Feb), NDVImon (Jun–Sep), NDVIdiff
Terrain (SRTM)	Elevation, Slope, Aspect

Land Cover Classes

ID	Class	Description
0	Water	Rivers, lakes, reservoirs
1	Forest	Dense / closed-canopy forest
2	Highland / Snow Cover	Alpine shrubland, seasonal snow, glaciers
3	Built-up Area	Urban areas, impervious surfaces
4	Barren Land	Bare rock, soil, scree

Training Data

Source	Count
Manual ground-truth points (GEE digitisation)	346
ESA WorldCover 2021 stratified samples	1,000
Total (after null removal)	1,337
Training split (70%, seed=42)	883
Validation split (30%)	425

Classifiers (GEE SMILE Library)

Classifier	Key Hyperparameters
Random Forest (RF)	200 trees, 3 vars/split, min leaf = 2, seed = 42
Gradient Tree Boosting (GTB)	200 trees, shrinkage = 0.05, seed = 42

---

📊 Results

Overall Accuracy

Classifier	Overall Accuracy	Cohen's Kappa	Macro AUC-ROC
Random Forest (RF)	75.76%	0.6956	0.9414
Gradient Tree Boosting (GTB)	76.00%	0.6990	0.9481

Both Kappa values exceed 0.60 — confirming substantial agreement beyond chance.

Per-class Metrics (%)

Class	RF F1	GTB F1	RF AUC	GTB AUC
Water	81.77	80.45	0.9526	0.9556
Forest	73.68	73.47	0.9370	0.9498
Highland / Snow Cover	66.18	66.19	0.9207	0.9194
Built-up Area	86.41	87.56	0.9756	0.9767
Barren Land	62.71	66.67	0.9209	0.9391

Built-up Area achieved the highest F1 (>86%) for both classifiers.
GTB outperforms RF by +3.96 F1 points for Barren Land.
RF achieves higher recall for Forest.
All five classes exceed AUC > 0.92 for both classifiers.

---

⚙️ Getting Started

Prerequisites

• A Google Earth Engine account
• Python 3.8+ with scikit-learn, matplotlib, numpy, pandas
• QGIS 3.x (for map visualisation)

Run the GEE Classification

1. Open GEE Code Editor
2. Copy the contents of Script.js and paste into the editor
3. Update the training asset path if needed:

   var trainingData = ee.FeatureCollection('projects/YOUR_PROJECT/assets/training_gangtok');

4. Click Run — the LULC map will appear in the map panel
5. Go to the Tasks tab and click Run to export the GeoTIFF to Drive

Run the Accuracy Assessment Notebook

git clone https://github.com/satwik-shreshth/LULC_Classification.git
cd LULC_Classification
pip install scikit-learn matplotlib numpy pandas
jupyter notebook LULC.ipynb

---

🛠️ Tools & Technologies

Tool	Purpose
Google Earth Engine	Cloud-based satellite data processing & classification
Sentinel-2 SR (ESA Copernicus)	Primary satellite imagery
USGS SRTM 30m DEM	Terrain features
ESA WorldCover 2021	Stratified training sample source
FAO GAUL 2015	District boundary
GEE SMILE Library	RF & GTB classifier training
scikit-learn	AUC-ROC, confusion matrices, evaluation plots
QGIS	Map cartography & export
Python / Jupyter	Analysis and visualisation

---

⚠️ Limitations

• Overall accuracy of ~76% reflects inherent spectral mixing in complex Himalayan terrain at 10-m resolution
• WorldCover-derived samples may propagate label noise in transition zones
• No post-classification spatial smoothing — salt-and-pepper noise visible in heterogeneous zones
• Persistent monsoon cloud cover may introduce temporal bias at high-altitude pixels
• Area estimates from pixel counts have not been bias-corrected for map accuracy

---

🔭 Future Work

• SAR-optical fusion — Sentinel-1 SAR for improved snow/bare rock discrimination
• Object-Based Image Analysis (OBIA) — reduce noise at class boundaries
• Deep learning — CNN/U-Net approaches exploiting spatial context
• Bias-corrected area estimation — following Olofsson et al. (2014)
• Temporal change detection — multi-year LULC change mapping

---

🙏 Acknowledgements

• India Space Academy (ISA) — Summer Training Programme 2025 opportunity and mentorship
• ESA — Sentinel-2 data (Copernicus Programme) & WorldCover 2021
• USGS — SRTM Digital Elevation Model
• Google LLC — Earth Engine cloud platform
• FAO — GAUL boundary dataset

---

Report submitted: 12 August 2025  |  India Space Academy, Department of Space Education

⭐ If you found this useful, please star the repository!