Mountain Green Cover Index (MGCI) Using Deep Learning

A Deep Learning Framework for Automated Vegetation Segmentation and SDG 15.4.2 Monitoring

This repository provides a complete, reproducible pipeline for calculating the Mountain Green Cover Index (MGCI) — UN Sustainable Development Goal Indicator 15.4.2 — using deep learning and satellite remote sensing. The framework is region-agnostic and can be applied to any mountainous region worldwide.

---

Table of Contents

• Overview
• Key Features
• Quick Start
• Repository Structure
• Detailed Pipeline
  • Phase 1: Data Export
  • Phase 2: Model Training
  • Phase 3: Inference & MGCI Calculation
• Using Pre-trained Weights
• Applying to Your Region
• Results
• Citation
• License
• Acknowledgments

---

Overview

The Mountain Green Cover Index (MGCI) measures the percentage of mountain area covered by green vegetation. It is a key indicator for monitoring progress toward SDG Target 15.4: "Ensure the conservation of mountain ecosystems."

Traditional vs. Deep Learning Approach

Aspect	Traditional (NDVI Threshold)	Our Approach (U-Net)
Method	Fixed NDVI > 0.3 threshold	Learned semantic segmentation
Inputs	NIR and Red bands only	6 bands (RGB, NIR, Elevation, Slope)
Context	Per-pixel, no spatial context	Spatial context via convolutions
Accuracy	Prone to false positives	Higher precision, fewer errors

What This Framework Provides

1. Automated Data Pipeline — Export satellite data for any region using Google Earth Engine
2. Deep Learning Model — U-Net with ResNet-50 encoder for vegetation segmentation
3. MGCI Calculation — Both planimetric and true surface area methods
4. Kapos Stratification — Results broken down by elevation class
5. Pre-trained Weights — Ready-to-use model for immediate inference

---

Key Features

• Region-Agnostic: Apply to any country or administrative region
• Fully Reproducible: Step-by-step notebooks with clear instructions
• Multiple MGCI Methods: Compare NDVI threshold vs. model predictions
• True Surface Area: Slope-corrected area calculations
• Kapos Classification: Results stratified by elevation class (300m to 4500m+)
• Publication-Ready Figures: Automated visualization generation
• Pre-trained Weights: Use our model or train your own

---

Quick Start

Option A: Use Pre-trained Weights (Fastest)

# 1. Clone the repository
git clone https://github.com/QEC-Team/MGCI-Deep-Learning.git
cd MGCI-Deep-Learning

# 2. Download pre-trained weights (see weights/README.md)

# 3. Open notebooks in Google Colab:
#    - Run Notebook 1 to export data for YOUR region
#    - Skip Notebook 2 (training)
#    - Run Notebook 3 for inference using pre-trained weights

Option B: Train Your Own Model

# 1. Clone the repository
git clone https://github.com/QEC-Team/MGCI-Deep-Learning.git
cd MGCI-Deep-Learning

# 2. Open notebooks in Google Colab:
#    - Run Notebook 1 to export data
#    - Run Notebook 2 to train the model
#    - Run Notebook 3 for inference

---

Repository Structure

MGCI-Deep-Learning/
│
├── README.md                          # This file
├── LICENSE                            # CC BY-NC-ND 4.0 License
├── CITATION.cff                       # Citation information
├── requirements.txt                   # Python dependencies
├── .gitignore                         # Git ignore rules
│
├── notebooks/
│   ├── MGCI_1_DataExport.ipynb       # Phase 1: Google Earth Engine export
│   ├── MGCI_2_Training.ipynb         # Phase 2: Model training
│   └── MGCI_3_Inference.ipynb        # Phase 3: Inference & MGCI calculation
│
├── weights/
│   └── README.md                      # Instructions for downloading weights
│
├── docs/
│   ├── SETUP.md                       # Detailed setup guide
│   ├── DATA_PREPARATION.md            # Region-specific data preparation
│   ├── METHODOLOGY.md                 # Technical methodology details
│   └── TROUBLESHOOTING.md             # Common issues and solutions
│
├── figures/
│   ├── pipeline_overview.png          # Pipeline diagram
│   ├── model_architecture.png         # U-Net architecture
│   ├── sample_results.png             # Example outputs
│   └── kapos_classification.png       # Elevation class diagram
│
└── examples/
    └── region_configs/                # Example configurations for regions
        ├── saudi_arabia_jazan.py
        └── template.py

---

Detailed Pipeline

Phase 1: Data Export

Notebook: MGCI_1_DataExport.ipynb

This notebook uses Google Earth Engine to export satellite data for your region of interest.

Data Sources

Data	Source	Resolution	Purpose
Optical Imagery	Sentinel-2 L2A	10m	RGB + NIR bands
Elevation	Copernicus GLO-30 DEM	30m (resampled to 10m)	Mountain classification
Slope	Derived from DEM	10m	Surface area correction
Administrative Boundaries	FAO GAUL	Vector	Region selection

Exported Bands

Index	Band	Source	Description
0	Blue	Sentinel-2 B2	Blue reflectance (0-10000)
1	Green	Sentinel-2 B3	Green reflectance (0-10000)
2	Red	Sentinel-2 B4	Red reflectance (0-10000)
3	NIR	Sentinel-2 B8	Near-infrared reflectance (0-10000)
4	Elevation	Copernicus DEM	Elevation in meters
5	Slope	Derived	Slope in degrees (0-90)
6	VegLabel	NDVI threshold	Binary vegetation label (0/1)

Key Parameters to Modify

# In the Config class:
COUNTRY = 'Saudi Arabia'           # Your country
REGION = 'Jazan'                   # Your region/province
YEAR = 2024                        # Year of analysis
CLOUD_COVER_MAX = 20               # Maximum cloud cover percentage
NDVI_THRESHOLD = 0.30              # Vegetation threshold
MOUNTAIN_THRESHOLD = 300           # Minimum elevation (meters)
PATCH_SIZE = 2560                  # Patch size in meters

Output

• GeoTIFF patches (7 bands each) saved to Google Drive
• Each patch covers ~2.56 × 2.56 km at 10m resolution (256 × 256 pixels)

---

Phase 2: Model Training

Notebook: MGCI_2_Training.ipynb

This notebook trains a U-Net model for vegetation segmentation.

Model Architecture

U-Net with ResNet-50 Encoder
├── Encoder: ResNet-50 (ImageNet pre-trained)
├── Decoder: U-Net decoder blocks
├── Input: 6 channels (Blue, Green, Red, NIR, Elevation, Slope)
├── Output: 1 channel (vegetation probability)
└── Parameters: ~32.5 million

Training Configuration

Parameter	Value	Description
Optimizer	AdamW	Adaptive optimizer with weight decay
Learning Rate	3×10⁻⁴	Initial learning rate
Weight Decay	1×10⁻⁴	L2 regularization
Batch Size	8	Samples per batch
Epochs	80	Maximum training epochs
Early Stopping	15	Patience (epochs without improvement)
Loss Function	0.5×Focal + 0.5×Dice	Combined loss for class imbalance
Scheduler	CosineAnnealingWarmRestarts	Learning rate schedule

Data Augmentation

Augmentation	Probability	Purpose
Horizontal Flip	0.5	Spatial invariance
Vertical Flip	0.5	Spatial invariance
Random Rotate 90°	0.5	Rotational invariance
Shift/Scale/Rotate	0.5	Geometric robustness
Gaussian Noise	0.3	Noise robustness
Brightness/Contrast	0.3	Radiometric robustness

Data Split

The data is split spatially (sorted by longitude) to prevent data leakage:

Split	Percentage	Purpose
Train	70%	Model training
Validation	15%	Hyperparameter tuning
Test	15%	Final evaluation

Output

• model_best.pth — Saved model weights
• Training history plots
• Confusion matrix
• Sample predictions

---

Phase 3: Inference & MGCI Calculation

Notebook: MGCI_3_Inference.ipynb

This notebook runs inference on all patches and calculates MGCI.

MGCI Formula

$$\text{MGCI} = \frac{\text{Green Mountain Area}}{\text{Total Mountain Area}} \times 100%$$

Methods Compared

Method	Description
NDVI Planimetric	Traditional NDVI > 0.3 threshold, flat area
NDVI True Surface	NDVI threshold with slope-corrected area
Model Planimetric	U-Net predictions, flat area
Model True Surface	U-Net predictions with slope-corrected area

True Surface Area Correction

For sloped terrain, the true surface area is larger than the planimetric (map) area:

$$\text{True Surface Area} = \frac{\text{Planimetric Area}}{\cos(\text{slope})}$$

Kapos Elevation Classes

Class	Name	Elevation Range
6	Foothills	300 – 1,000 m
5	Lower Montane	1,000 – 1,500 m
4	Montane	1,500 – 2,500 m
3	Upper Montane	2,500 – 3,500 m
2	Alpine	3,500 – 4,500 m
1	Nival	> 4,500 m

Output

• mgci_results.json — Complete MGCI results
• mgci_by_kapos_class.json — Results by elevation class
• Publication-ready figures

---

Using Pre-trained Weights

We provide pre-trained weights that can be used directly for inference on new regions.

Download Weights

See weights/README.md for download instructions.

Model Details

Property	Value
Architecture	U-Net + ResNet-50
Input Channels	6
Training Region	Jazan, Saudi Arabia
Training Patches	878
Best Epoch	68
Validation IoU	0.8475
Test Dice	0.9263

Using Pre-trained Weights

In Notebook 3, simply point to the downloaded weights:

Config.MODEL_PATH = '/path/to/model_best.pth'

The model will load and run inference on your exported patches.

---

Applying to Your Region

Step-by-Step Guide

1. Choose Your Region

Select any administrative region available in the FAO GAUL database:

• Country level (e.g., "Switzerland")
• Province/State level (e.g., "Colorado")
• District level (e.g., "Zermatt")

2. Modify Notebook 1 Configuration

class Config:
    # Change these for your region
    COUNTRY = 'Your Country'
    REGION = 'Your Region'
    
    # Adjust based on your region's characteristics
    YEAR = 2024
    CLOUD_COVER_MAX = 20
    NDVI_THRESHOLD = 0.30
    MOUNTAIN_THRESHOLD = 300  # Kapos minimum
    
    # Adjust date range for your region's growing season
    START_DATE = '2024-01-01'
    END_DATE = '2024-12-31'

3. Run the Pipeline

Step	Notebook	Time (Approx.)
1	Export Data	12-48 hours (depends on region size)
2	Train Model	1-2 hours (GPU recommended)
3	Run Inference	5-10 minutes

4. Interpret Results

The output JSON files contain:

• Total mountain area (km²)
• Green mountain area (km²)
• MGCI values for each method
• Results stratified by Kapos class

Regional Considerations

Region Type	Considerations
Tropical	Adjust date range to avoid monsoon; may need lower cloud threshold
Arid	NDVI threshold may need adjustment; sparse vegetation
Alpine	High elevation classes will be populated
Temperate	Consider seasonal variation; growing season dates important

---

Results

Model Performance (Test Set)

Metric	Value
Accuracy	96.57%
Precision	92.06%
Recall	93.21%
F1 Score	92.63%
IoU	86.28%
Dice	92.63%

Training History

MGCI Results (Jazan, Saudi Arabia)

Method	MGCI (%)
NDVI (Planimetric)	21.76%
NDVI (True Surface)	22.28%
Model (Planimetric)	18.88%
Model (True Surface)	19.36%

Area Breakdown

MGCI by Elevation Class

Key Findings

• Model predictions are more conservative than NDVI threshold
• True surface area correction adds ~0.5% to MGCI values
• Model reduces false positives in bare soil and rocky areas

---

Citation

If you use this code or methodology in your research, please cite:

@software{mgci_deep_learning_2024,
  author = {Alorf, Abdulaziz and Alsultan, Muhannad and Alghonaim, Thamer and Alwazzan, Bandar},
  title = {MGCI-Deep-Learning: A Framework for Mountain Green Cover Index Calculation Using Deep Learning},
  year = {2024},
  publisher = {GitHub},
  url = {https://github.com/QEC-Team/MGCI-Deep-Learning}
}

Publication (In Preparation):

Alorf, A., Alsultan, M., Alghonaim, T., & Alwazzan, B. (2024). Developing a National Framework for Estimating Mountain Green Cover Using Deep Learning and Satellite Remote Sensing. Manuscript in preparation.

---

License

This project is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License (CC BY-NC-ND 4.0).

You are free to:

• Share — Copy and redistribute the material in any medium or format

Under the following terms:

• Attribution — You must give appropriate credit and cite this work
• NonCommercial — You may not use the material for commercial purposes
• NoDerivatives — You may not distribute modified versions

See LICENSE for the full license text.

---

Acknowledgments

• Google Earth Engine — Satellite data access and processing
• European Space Agency — Sentinel-2 imagery
• Copernicus Programme — GLO-30 DEM
• FAO — GAUL administrative boundaries and MGCI methodology

---

Contact

For questions or collaboration inquiries, please open an issue on this repository.

---

MGCI Deep Learning Framework