This project aims to detect deforestation from satellite images using a U-Net deep learning model. It leverages automated satellite data collection, image segmentation techniques, and Earth Engine + TensorFlow pipelines to help environmental organizations monitor illegal deforestation in near real-time.

Can an automated model built on satellite imagery help environmental organizations monitor illegal deforestation in near real-time?

We aim to answer this by building a robust deep learning pipeline that:

• Collects satellite imagery for selected regions and years.
• Trains a U-Net segmentation model on forest/non-forest (or deforested) labels.
• Allows users to input coordinates + year and visualize predictions on a Leaflet map.

We use Google Earth Engine (GEE) to fetch Sentinel-2 or Landsat-8 imagery and NDVI indices.

• Imagery: Landsat-8 (30m resolution), optionally Sentinel-2 (10m resolution).
• Labels: Derived using NDVI and/or forest loss layers from Global Forest Change dataset.

• Region-wise time-lapse collection.
• NDVI-based labeling: Thresholding NDVI to identify vegetation loss.
• Tiling large images into patches.
• Saving as .npy and .tfrecord formats for model training.

The model uses a U-Net architecture implemented in TensorFlow, designed for image segmentation tasks.

• Input: Satellite image patches (RGB or NDVI composite).
• Output: Binary mask indicating deforested areas.

• Batch normalization and dropout regularization
• Dice coefficient + binary cross-entropy loss
• Data augmentation (rotation, flipping, contrast)

python train.py --epochs 50 --batch_size 16 --data_dir ./data --save_model ./models/unet_deforestation.h5

The trained model is evaluated on a held-out test set using the following metrics:

• IoU (Intersection over Union): Measures the overlap between predicted deforested area and ground truth.
• Dice Coefficient: Especially useful for imbalanced classes.
• Precision & Recall: To understand false positives and false negatives.
• F1 Score: Harmonic mean of precision and recall.

Visit: Deforestation Detection on Hugging Face

Features:

• Interactive Leaflet map to select coordinates
• Year input to analyze temporal deforestation
• On-click image fetching and prediction overlay
• Downloadable prediction masks for offline use

The backend takes user inputs (lat, lon, year), queries GEE for imagery, preprocesses it into patches, passes it to the trained U-Net model, and returns a combined prediction image with overlays.

git clone [https://github.com/Ojas-Rohatgi/Deforestation-Detection](https://github.com/Ojas-Rohatgi/Deforestation-Detection)
cd Deforestation-Detection

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt

Ensure you have satellite imagery and labeled NDVI masks stored in the data/ folder.

python training.py --epochs 50 --batch\_size 16 --data\_dir ./data --save\_model ./model/unet\_deforestation.h5

You can configure the number of epochs, batch size, and save path as needed.

cd app
python app.py

This will launch a web interface where you can:

• Select a region using the interactive Leaflet map
• Input a year (e.g., 2020)
• Fetch satellite imagery and get prediction overlays

This app is live at: 🔗 Hugging Face Spaces – ojasrohatgi/Deforestation-Detection

• 🌍 Interactive map for selecting location
• 📅 Year input for time-based analysis
• 📷 Automatic image fetch via GEE
• 🧠 Real-time inference with trained U-Net
• 🖼️ Visual overlay of predictions
• ⬇️ Option to download the prediction mask

• Multi-year change detection (e.g., forest loss between 2015–2023)
• Visual heatmaps of deforestation severity
• Region-specific fine-tuning for Southeast Asia, Amazon, and Africa
• Support for multispectral and SAR imagery
• Webhooks/API integration for automated NGO alerts

Ojas Rohatgi Final Year B.Tech – Computer Science (Data Science & AI) SRM University, Sonepat, Haryana, India

• 🔗 GitHub: Ojas-Rohatgi
• 🔗 Hugging Face: ojasrohatgi

This project is licensed under the Apache License. See the LICENSE file for full license text.

If you found this useful:

• 🌟 Star this repository on GitHub
• 🚀 Share the Hugging Face app with your network
• 📢 Report issues or suggest improvements in the Issues tab