lloydflorens/mauritius-landcover

Mauritius Land Cover Classification

Deep learning-based land cover classification for Mauritius using Sentinel-2 satellite imagery and U-Net with ResNet50 encoder.

Overview

This project uses a U-Net segmentation model with a ResNet50 encoder to classify land cover types across Mauritius from Sentinel-2 multispectral satellite imagery.

Land Cover Classes (Apple Maps Style)

Model Performance

Features

• —Live Interactive Map: Real-time land cover classification as you pan across Mauritius
• —9-Band Input: Uses B2, B3, B4, B8, B11, B12 + NDVI, NDWI, NDBI indices
• —Apple Maps-inspired styling: Clean, aesthetic color palette
• —Google Earth Engine integration: Automatic Sentinel-2 imagery download

Installation

# Clone the repository
git clone https://github.com/yourusername/mauritius-landcover.git
cd mauritius-landcover

# Create virtual environment
python -m venv venv
source venv/bin/activate  # Linux/Mac
# or: venv\Scripts\activate  # Windows

# Install dependencies
pip install -r requirements.txt

# Authenticate with Google Earth Engine
earthengine authenticate

Usage

Live Interactive Map

python src/web/live_interactive_map.py
# Open http://localhost:5003 in your browser

Pan around the map to automatically fetch Sentinel-2 imagery and classify land cover in real-time.

Training Pipeline

# 1. Download training tiles from Google Earth Engine
python scripts/gather_targeted_samples.py

# 2. Auto-label tiles using spectral indices
python scripts/auto_label_tiles_fixed.py

# 3. Train model with class weights
python scripts/train_enhanced.py --epochs 50 --use-class-weights

Project Structure

mauritius-landcover/
├── src/
│   ├── web/
│   │   └── live_interactive_map.py    # Flask web app with Leaflet.js
│   ├── models/
│   │   └── unet.py                     # U-Net model definition
│   └── data/
│       └── dataset.py                  # PyTorch dataset class
├── scripts/
│   ├── gather_targeted_samples.py      # Download training tiles from GEE
│   ├── auto_label_tiles_fixed.py       # Rule-based auto-labeling
│   ├── train_enhanced.py               # Training script with class weights
│   └── create_training_dataset.py      # Dataset creation utilities
├── checkpoints/                         # Model weights (not in repo)
├── data/
│   └── training/tiles/                  # Training data (not in repo)
└── requirements.txt

Technical Details

Input Data

• —6 Spectral Bands: B2 (Blue), B3 (Green), B4 (Red), B8 (NIR), B11 (SWIR1), B12 (SWIR2)
• —3 Spectral Indices: NDVI, NDWI, NDBI
• —Resolution: 10m per pixel
• —Tile Size: 256 x 256 pixels

Model Architecture

• —Encoder: ResNet50 (pretrained on ImageNet)
• —Decoder: U-Net style upsampling
• —Input Channels: 9 (modified first conv layer)
• —Output Classes: 7 (including background)

Training

• —Loss: Class-weighted CrossEntropyLoss
• —Optimizer: AdamW with discriminative learning rates
• —Augmentation: Random flips, rotations, color jittering
• —Epochs: 50

Requirements

• —Python 3.8+
• —PyTorch 2.0+
• —segmentation-models-pytorch
• —Google Earth Engine Python API
• —Flask
• —rasterio
• —numpy, Pillow

License

MIT License

Author

Lloyd Florens - PhD Research, Mauritius Land Cover Classification